2nd detection / segmentation challengesegmentation leaderboard (i) coco ap (over all iou) coco ap...

109
2 nd Detection / Segmentation Challenge Yin Cui, Tsung-Yi Lin, Matteo Ruggero Ronchi, Genevieve Patterson ImageNet and COCO Visual Recognition Challenges Workshop Sunday, October 9th, ECCV 2016

Upload: others

Post on 25-Sep-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

2ndDetection / Segmentation Challenge

Yin Cui, Tsung-Yi Lin, Matteo Ruggero Ronchi, Genevieve Patterson

ImageNet and COCO Visual Recognition Challenges WorkshopSunday, October 9th, ECCV 2016

Page 2: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 542

Workshop Organizers

Yin CuiCornell Tech

Matteo Ruggero RonchiCaltech

Genevieve PattersonBrown University

Michael MaireSerge BelongieLubomir BourdevRoss GirshickJames HaysPietro PeronaLarry ZitnickPiotr Dollár

Workshop Advisors:Deva RamananPietro PeronaMichael MaireLubomir BourdevSerge BelongieMatteo Ruggero RonchiGenevieve PattersonYin Cui

Award Committee:

Tsung-Yi LinCornell Tech

Page 3: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 543

Outline

1. Download MS COCO Train / Val set

Participate in challenge?

Yes

No 3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

2. Develop the algorithm

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 4: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 543

Outline

1. Download MS COCO Train / Val set

Participate in challenge?

Yes

No 3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

2. Develop the algorithm

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 5: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 544

• 80 object categories • 200k images• 1.2M instances (350k people)• Every instance segmented

Available for download atmscoco.org

COCO Dataset

Page 6: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 544

• 80 object categories • 200k images• 1.2M instances (350k people)• Every instance segmented

Available for download atmscoco.org

COCO Dataset

• 106k people with keypoints

Page 7: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 545

Available for download atmscoco.org/external

COCO 3rd Party Datasets

Page 8: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 546

1. Download MS COCO Train / Val set

Participate in challenge?

Yes

No

2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 9: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 546

1. Download MS COCO Train / Val set

Participate in challenge?

Yes

No

2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 10: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 547

Shout-out to previous algorithms!

Page 11: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 547

Shout-out to previous algorithms!

Page 12: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 548

1. Download MS COCO Train / Val set

Participate in challenge?

Yes

No

2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 13: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 548

1. Download MS COCO Train / Val set

Participate in challenge?

Yes

No

2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 14: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 549

Challenges at ECCV 2016

Page 15: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5410

Participate in challenge?

Yes

No

1. Download MS COCO Train / Val set 2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 16: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5410

Participate in challenge?

Yes

No

1. Download MS COCO Train / Val set 2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 17: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5411

MS COCO Test Sets

Page 18: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5411

The 2015/2016 MS COCO Test set consists of ~80k test images.

MS COCO Test Sets

Page 19: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5411

The 2015/2016 MS COCO Test set consists of ~80k test images.

Test-dev (development) Debugging, Validation and Ablation Studies. Allows unlimited submission to the evaluation server.

MS COCO Test Sets

Page 20: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5411

The 2015/2016 MS COCO Test set consists of ~80k test images.

Test-dev (development) Debugging, Validation and Ablation Studies. Allows unlimited submission to the evaluation server.

Test-standard (publications) Used to score entries for the Public Leaderboard.

MS COCO Test Sets

Page 21: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5411

The 2015/2016 MS COCO Test set consists of ~80k test images.

Test-dev (development) Debugging, Validation and Ablation Studies. Allows unlimited submission to the evaluation server.

Test-standard (publications) Used to score entries for the Public Leaderboard.

Test-challenge (competitions) Used to score workshop competition.

MS COCO Test Sets

Page 22: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5411

The 2015/2016 MS COCO Test set consists of ~80k test images.

Test-dev (development) Debugging, Validation and Ablation Studies. Allows unlimited submission to the evaluation server.

Test-standard (publications) Used to score entries for the Public Leaderboard.

Test-challenge (competitions) Used to score workshop competition.

Test-reserve (security) Used to estimate overfitting. Scores on this set are never released.

MS COCO Test Sets

Page 23: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5412

Participate in challenge?

Yes

No

1. Download MS COCO Train / Val set 2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 24: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5412

Participate in challenge?

Yes

No

1. Download MS COCO Train / Val set 2. Develop the algorithm

Outline

3. Download MS COCO Test-Dev

3. Download MS COCO Test-Full

4. Upload to CodaLab (unlimited)

4. Upload to CodaLab (5 times max)

Page 25: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5413

Evaluation Server Usage

Submissions to all test sets

Page 26: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5414

Evaluation Metrics

Page 27: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5415

Evaluation Metrics

• AP is averaged over multiple IoU values between 0.5 and 0.95.

Challenges Score: AP• More comprehensive metric than

the traditional AP at a fixed IoU value (0.5 for PASCAL).

Page 28: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 54

• AP is averaged over instance size:• small (A < 32 x 32)• medium (32x 32 < A < 96 x 96)• large (A > 96 x 96)

16

Evaluation Metrics

A<32x32

32x32<A<96x96

A>96x96Other Scores: Size AP

Page 29: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5417

Evaluation Metrics

Other Scores: AR

• Measures the maximum recall over a fixed number of detections allowed in the image of 1, 10, 100.

• AR is averaged over small (A < 32 x 32), medium (32x 32 < A < 96 x 96) and large (A > 96 x 96) instances of objects.

Page 30: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5418

Evaluation Ambiguity

Which one is better?

Page 31: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5418

Evaluation Ambiguity

Which one is better?

Page 32: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5418

Evaluation Ambiguity

Which one is better?

Ground-Truth BBox

Page 33: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5418

Evaluation Ambiguity

Which one is better?

Detection BBoxGround-Truth BBox

Page 34: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5418

Evaluation Ambiguity

IoU = 0.5 IoU = 0.7 IoU = 0.95

Which one is better?

Detection BBoxGround-Truth BBox

Page 35: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5419

COCO Challenges Results

Page 36: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5420

Bounding Boxes Leaderboard (I)

COCO AP (over all IoU)

Page 37: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5420

0%

10%

20%

30%

40%

50%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

Fast

R-CNN*

(VGG-16)

Bounding Boxes Leaderboard (I)

COCO AP (over all IoU)

Page 38: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5420

0%

10%

20%

30%

40%

50%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

Fast

R-CNN*

(VGG-16)

Bounding Boxes Leaderboard (I)

COCO AP (over all IoU)

(*) Performance on Test-Dev

Page 39: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5420

0%

10%

20%

30%

40%

50%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

Fast

R-CNN*

(VGG-16)

Bounding Boxes Leaderboard (I)

COCO AP (over all IoU)

(*) Performance on Test-Dev

Page 40: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5420

0%

10%

20%

30%

40%

50%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

Fast

R-CNN*

(VGG-16)

Bounding Boxes Leaderboard (I)

COCO AP (over all IoU)

(*) Performance on Test-Dev (**) 2015 Winner

Page 41: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5420

0%

10%

20%

30%

40%

50%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

Fast

R-CNN*

(VGG-16)

Bounding Boxes Leaderboard (I)

COCO AP (over all IoU)

(*) Performance on Test-Dev

+22% absolute+110% relative

(**) 2015 Winner

Page 42: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5420

0%

10%

20%

30%

40%

50%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

Fast

R-CNN*

(VGG-16)

Bounding Boxes Leaderboard (I)

COCO AP (over all IoU)

(*) Performance on Test-Dev

+22% absolute+110% relative

(**) 2015 Winner

+4.2% absolute+11.2% relative

Page 43: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5421

Segmentation Leaderboard (I)

COCO AP (over all IoU)

Page 44: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5421

0%

10%

20%

30%

40%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

Segmentation Leaderboard (I)

COCO AP (over all IoU)

Page 45: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5421

0%

10%

20%

30%

40%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

Segmentation Leaderboard (I)

COCO AP (over all IoU)

(**) 2015 Winner

Page 46: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5421

0%

10%

20%

30%

40%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

Segmentation Leaderboard (I)

COCO AP (over all IoU)

(**) 2015 Winner

Page 47: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5421

0%

10%

20%

30%

40%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

Segmentation Leaderboard (I)

COCO AP (over all IoU)

(**) 2015 Winner

Page 48: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5421

0%

10%

20%

30%

40%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

Segmentation Leaderboard (I)

COCO AP (over all IoU)

(**) 2015 Winner

+9.1% absolute+32.3% relative

Page 49: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5421

0%

10%

20%

30%

40%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

Segmentation Leaderboard (I)

COCO AP (over all IoU)

COCO AP for segmentation winner trails the one for bbox detection by ~4%:

• Last year the gap was ~10%• Localization is harder

for segmentation

(**) 2015 Winner

+9.1% absolute+32.3% relative

Page 50: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5422

Bounding Boxes Leaderboard (II)

Object Localization is improving

Page 51: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5422

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

AP_50 AP_75

Bounding Boxes Leaderboard (II)

Object Localization is improving

Page 52: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5422

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

AP_50 AP_75

Bounding Boxes Leaderboard (II)

Object Localization is improving

Page 53: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5422

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

AP_50 AP_75

Bounding Boxes Leaderboard (II)

Object Localization is improving

objects correctly detected but not well localized17% AP

Page 54: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5422

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRAVC**

Trimps-S

oush

en

Imag

ine La

b

Cmu-a2-v

gg16

ToCon

coctP

ellucid Wall

hust-

mclab

AP_50 AP_75

Bounding Boxes Leaderboard (II)

Object Localization is improving

17% AP 19% AP

Page 55: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5423

Segmentation Leaderboard (II)

Mask localization can improve

Page 56: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5423

0%

10%

20%

30%

40%

50%

60%

70%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

AP_50 AP_75

Segmentation Leaderboard (II)

Mask localization can improve

Page 57: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5423

0%

10%

20%

30%

40%

50%

60%

70%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

AP_50 AP_75

Segmentation Leaderboard (II)

Mask localization can improve

Page 58: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5423

0%

10%

20%

30%

40%

50%

60%

70%

MSRAG-R

MI

MSRAVC**

anon

ymou

s

AP_50 AP_75

Segmentation Leaderboard (II)

Mask localization can improve

20% AP

Page 59: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5424

Bounding Boxes vs Segmentation

Segmentation provides great bounding boxes!

Page 60: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5424

Bounding Boxes vs Segmentation

Segmentation provides great bounding boxes!

(*) 2015 Winner

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRA (seg

m)

MSRAVC (bbox

)*

AP AP_75 AP_50

Page 61: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5424

Bounding Boxes vs Segmentation

Segmentation provides great bounding boxes!

COCO AP for segmentation winner trails the one for bbox detection by ~2%:

(*) 2015 Winner

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRA (seg

m)

MSRAVC (bbox

)*

AP AP_75 AP_50

Page 62: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5424

Bounding Boxes vs Segmentation

Segmentation provides great bounding boxes!

COCO AP for segmentation winner trails the one for bbox detection by ~2%:

(*) 2015 Winner

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRA (seg

m)

MSRAVC (bbox

)*

AP AP_75 AP_50

Page 63: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5424

Bounding Boxes vs Segmentation

Segmentation provides great bounding boxes!

COCO AP for segmentation winner trails the one for bbox detection by ~2%:

• Results in 2nd place in the bbox challenge!

(*) 2015 Winner

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRA (seg

m)

MSRAVC (bbox

)*

AP AP_75 AP_50

Page 64: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5424

Bounding Boxes vs Segmentation

Segmentation provides great bounding boxes!

COCO AP for segmentation winner trails the one for bbox detection by ~2%:

• Results in 2nd place in the bbox challenge!

• Gap is about constant at multiple IoU values.

(*) 2015 Winner

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRA (seg

m)

MSRAVC (bbox

)*

AP AP_75 AP_50

Page 65: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5424

Bounding Boxes vs Segmentation

Segmentation provides great bounding boxes!

COCO AP for segmentation winner trails the one for bbox detection by ~2%:

• Results in 2nd place in the bbox challenge!

• Gap is about constant at multiple IoU values.

• Participate in Segmentation Challenge!

(*) 2015 Winner

0%

10%

20%

30%

40%

50%

60%

70%

G-RMI

MSRA (seg

m)

MSRAVC (bbox

)*

AP AP_75 AP_50

Page 66: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5425

Performance Breakdown (I)

COCO AP varies across supercategories and size

Page 67: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5425

0%

10%

20%

30%

40%

50%

60%

anim

alou

tdoo

rve

hicle

elect

roni

cpe

rson

appl

iance

furn

iture

spor

tsfo

odin

door

kitch

enac

cess

ory

Performance Breakdown (I)

COCO AP varies across supercategories and size

Page 68: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5425

0%

10%

20%

30%

40%

50%

60%

anim

alou

tdoo

rve

hicle

elect

roni

cpe

rson

appl

iance

furn

iture

spor

tsfo

odin

door

kitch

enac

cess

ory

Performance Breakdown (I)

COCO AP varies across supercategories and size

Page 69: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5425

0%

10%

20%

30%

40%

50%

60%

anim

alou

tdoo

rve

hicle

elect

roni

cpe

rson

appl

iance

furn

iture

spor

tsfo

odin

door

kitch

enac

cess

ory

Performance Breakdown (I)

COCO AP varies across supercategories and size

Performance across teams improved on all supercategories

• Average AP increase of ~10%.• Average Standard Deviation

decrease of ~1%.

Page 70: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5426

Performance Breakdown (II)

Impact of size on performance

0%

15%

30%

45%

60%

2015 2016

small

medium large

Page 71: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5426

Performance Breakdown (II)

Impact of size on performance

0%

15%

30%

45%

60%

2015 2016

small

medium large

+33%

Page 72: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5426

Performance Breakdown (II)

Impact of size on performance

0%

15%

30%

45%

60%

2015 2016

small

medium large

+33%

+53%

Page 73: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5426

Performance Breakdown (II)

Impact of size on performance

0%

15%

30%

45%

60%

2015 2016

small

medium large

+33%

+53%

+118%!!

Page 74: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

Page 75: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

Page 76: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

(*) 2015 Winner

Page 77: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

0 % 80%AP

0 %

80 %

AP

(*) 2015 Winner

Page 78: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

0 % 80%AP

0 %

80 %

AP

(*) 2015 Winner

Page 79: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

0 % 80%AP

0 %

80 %

AP

R2 = 0.98

(*) 2015 Winner

Page 80: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

0 % 80%AP

0 %

80 %

AP

R2 = 0.98

(*) 2015 Winner

Segmentation

G-RMI

MSR

A

Page 81: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

0 % 80%AP

0 %

80 %

AP

0 % 80%AP

R2 = 0.98

(*) 2015 Winner

Segmentation

G-RMI

MSR

A

Page 82: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

0 % 80%AP

0 %

80 %

AP

0 % 80%AP

R2 = 0.98

(*) 2015 Winner

Segmentation

G-RMI

MSR

A

Page 83: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5427

Correlation between methods

How similarly do algorithms perform?

G-R

MI

MSRAVC*

Bounding Boxes

0 % 80%AP

0 %

80 %

AP

0 % 80%AP

R2 = 0.98 R2 = 0.97

(*) 2015 Winner

Segmentation

G-RMI

MSR

A

Page 84: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5428

Bounding Box Detection Errors

Page 85: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5428

Bounding Box Detection Errors

How similarly do top algorithms perform?

Page 86: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5428

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

Bounding Box Detection Errors

How similarly do top algorithms perform?

0 0.2 0.4 0.6 0.8 1recall

0

0.2

0.4

0.6

0.8

1

prec

ision

overall-all-all

[.456] C75[.623] C50[.686] Loc[.700] Sim[.723] Oth[.925] BG[1.00] FN

G-RMI

Page 87: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5428

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

Bounding Box Detection Errors

How similarly do top algorithms perform?

0 0.2 0.4 0.6 0.8 1recall

0

0.2

0.4

0.6

0.8

1

prec

ision

overall-all-all

[.456] C75[.623] C50[.686] Loc[.700] Sim[.723] Oth[.925] BG[1.00] FN

G-RMI MSRAVC*

0 0.2 0.4 0.6 0.8 1recall

0

0.2

0.4

0.6

0.8

1

prec

ision

overall-all-all

[.399] C75[.589] C50[.682] Loc[.695] Sim[.713] Oth[.870] BG[1.00] FN

(*) 2015 Winner

Page 88: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5429

Bounding Box Detection Errors (I)

Page 89: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5429

Bounding Box Detection Errors (I)

What type of errors are algorithms doing?

Page 90: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5429

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

Bounding Box Detection Errors (I)

Page 91: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5429

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

Bounding Box Detection Errors (I)

G-RMI MSRAVC*

(*) 2015 Winner

Page 92: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5429

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

0 0.2 0.4 0.6 0.8 1recall

0

0.2

0.4

0.6

0.8

1

prec

ision

person-person-all

[.582] C75[.812] C50[.875] Loc[.875] Sim[.886] Oth[.970] BG[1.00] FN

0 0.2 0.4 0.6 0.8 1recall

0

0.2

0.4

0.6

0.8

1

prec

ision

person-person-all

[.510] C75[.724] C50[.832] Loc[.832] Sim[.841] Oth[.911] BG[1.00] FN

Bounding Box Detection Errors (I)

G-RMI MSRAVC*

(*) 2015 Winner

Page 93: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5430

Bounding Box Detection Errors (II)

Page 94: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5430

Bounding Box Detection Errors (II)

What type of errors are algorithms doing?

Page 95: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5430

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

Bounding Box Detection Errors (II)

Page 96: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5430

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

Bounding Box Detection Errors (II)

G-RMI MSRAVC*

(*) 2015 Winner

Page 97: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5430

0 0.2 0.4 0.6 0.8 1recall

0

0.2

0.4

0.6

0.8

1

prec

ision

overall-all-small

[.244] C75[.416] C50[.506] Loc[.518] Sim[.533] Oth[.824] BG[1.00] FN

0 0.2 0.4 0.6 0.8 1recall

0

0.2

0.4

0.6

0.8

1

prec

ision

overall-all-small

[.175] C75[.343] C50[.469] Loc[.476] Sim[.484] Oth[.709] BG[1.00] FN

AP @ IoU = [0.5; 0.75]

AP @ IoU = 0.1

Super-category FP removed

Category FP removed

Background FP removed

All errors are removed

Bounding Box Detection Errors (II)

G-RMI MSRAVC*

(*) 2015 Winner

Page 98: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

Page 99: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

2016 Detection Challenge Take-aways

Page 100: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

• MSRAVC 2015 set a very high bar for performance.

2016 Detection Challenge Take-aways

Page 101: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

• MSRAVC 2015 set a very high bar for performance.• G-RMI imroved COCO AP by 4% absolute, 11% relative.

2016 Detection Challenge Take-aways

Page 102: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

• MSRAVC 2015 set a very high bar for performance.• G-RMI imroved COCO AP by 4% absolute, 11% relative.• MSRA 2016 segmentation algorithm is great on bboxes.

2016 Detection Challenge Take-aways

Page 103: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

• MSRAVC 2015 set a very high bar for performance.• G-RMI imroved COCO AP by 4% absolute, 11% relative.• MSRA 2016 segmentation algorithm is great on bboxes.• Performance on all classes has improved across entries.

2016 Detection Challenge Take-aways

Page 104: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

• MSRAVC 2015 set a very high bar for performance.• G-RMI imroved COCO AP by 4% absolute, 11% relative.• MSRA 2016 segmentation algorithm is great on bboxes.• Performance on all classes has improved across entries.• Localization improved greatly in both challenges.

2016 Detection Challenge Take-aways

Page 105: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

• MSRAVC 2015 set a very high bar for performance.• G-RMI imroved COCO AP by 4% absolute, 11% relative.• MSRA 2016 segmentation algorithm is great on bboxes.• Performance on all classes has improved across entries.• Localization improved greatly in both challenges.• High relative improvement on small object instances.

2016 Detection Challenge Take-aways

Page 106: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5431

Summary of Findings

• MSRAVC 2015 set a very high bar for performance.• G-RMI imroved COCO AP by 4% absolute, 11% relative.• MSRA 2016 segmentation algorithm is great on bboxes.• Performance on all classes has improved across entries.• Localization improved greatly in both challenges.• High relative improvement on small object instances.• False negatives are reduced, thus better recall of teams.

2016 Detection Challenge Take-aways

Page 107: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 5432

Challenges Ranking

Page 108: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 54

G-RMI 1st 2nd

MSRA - 1st

Trimps-Soushen 2nd -

Imagine Lab 3rd -

UofA 5th -

1026 - 3rd

32

Challenges RankingTeam BBox Segmentation

Page 109: 2nd Detection / Segmentation ChallengeSegmentation Leaderboard (I) COCO AP (over all IoU) COCO AP for segmentation winner trails the one for bbox detection by ~4%: • Last year the

/ 54

G-RMI 1st 2nd

MSRA - 1st

Trimps-Soushen 2nd -

Imagine Lab 3rd -

UofA 5th -

1026 - 3rd

32

Challenges Ranking

Invited Speakers:• G-RMI / Object Detection / (2:30pm - 2:45pm)• MSRA / Segmentation / (2:45pm - 3:00pm)

Team BBox Segmentation