saliency-assisted navigation of very large landscape images cheuk yiu ipamitabh varshney
TRANSCRIPT
Saliency-Assisted Navigation of Very Large Landscape Images
Cheuk Yiu Ip Amitabh Varshney
Very Large Landscape Images
• Stitch images to create multi-gigapixel very large images • But WHERE should we start looking?
• Image Acquisition:• Gigapan• MS HDView
ToG 2007• Image Stitching:
• Kazhdan et al ToG 2008, 2010
• Summa et al ToG 2010
Visual Knowledge Discovery
• Visual knowledge discovery• Identify what is interesting• Visualize them
Challenges Contributions
Visual Scalability: Sliding-Window Saliency
Information Scalability: Anomaly Detection
Data Scalability: Parallel filtering, Saliency Storage
Validation: Validate against Web Community Tags
Results Preview
Visual Scalability
Information Scalability
• Design effective algorithms to process large images• The SMALL unique regions in the large images contain the
MOST information• Identify informative regions from repetitive scene elements
Data Scalability
• Very large images represent a large amount of data 5 Gpix RGBA = 20GB uncompressed
• Multicore and manycore parallel processing• Requires efficient algorithms O(n) and out-of-core GPU methods
Overview
• Sliding-Window Saliency Map• Detection Anomalous Regions• Interactive Exploration
Traditional Multiscale Image Saliency
• Detects “Pop-out” spots from the scene• Inspired by human visual system• Pre-attentive vision
• Find multiscale contrasting regions • Intensity, Color Opponencies (I, RG, BY)
1. Convolve (I, RG, BY) with Difference of Gaussians (DoG) filter (σ is stdev)
2. Repeat on downsampled images for multiscales
Image Saliency• Itti et al.
PAMI,1998• Bruce et al.
IJCV,2009• Goferman et al.
CVPR 2010• Work on small
images, very accurate but slow.
Multiscale Aggregation
• Works well on small images• If we have many more scales …• Large regions dominate small regions• Wait… we don’t want to miss the small
regions
• Traditional multiscale saliency is insufficient
Our Sliding-Window Aggregation
• We see different things at different zoom levels
• One saliency map per level• Only aggregate up to 4x• Use a sliding-window across
scales• Why 4x? • Eye resolution difference ~5x
All (σ - 256σ)
σ – 4σ
4σ – 16σ
16σ – 64σ
There are still too many regions…
• 18,000+ regions in 1.3Gpix (5 hours if a user spends 1s on each)
• Regions are enlarged for visibility• There are many contrasting repetitive elements
Information Discovery
• Identify the informative regions from the salient regions• Compare regions to find the most different ones• Detect the anomalous regions and outliers
• Visual Data Analysis• Mesh and Volume Saliency (Lee et al. ToG 2005, Kim et al. TVCG 2006) • Video Summarization (Daniel et al. Vis 2003)• Flow and Information Theory (Janicke et al. TVCG 2010)• Molecular Dynamics Layout (Patro et al. Biovis 2011)
Image Region Descriptors
• Represent salient regions by histograms (rotational invariance)• Global Colors RGB, HSV, CIELAB: Not discriminative• Local Edges: Too discriminative • Histograms of colors in 8x8 moving windows work
well(MPEG-7 CSD)• Compare histograms, p, q, by the Euclidean distance
• Uniqueness, U(p), is the average distance of p to its k-Nearest-Neighbors.
• Repeating regions have a low U(p)• Distinct regions have a high U(p)• Spatial data structures (kD-trees) accelerate the retrieval
k-Nearest-Neighbors Anomaly Detection
Where are they … ?
• Top 3% (500) of the most distinct regions.• Most of the repeating region are eliminated.• Can you see the remaining regions?
Visualizing the Detected Regions
• Problem: Small regions of interests are NOT visible
• Adaptively enlarge regions• Determine the scale and
colors by the region’s rank of uniqueness
• Increase when zooming out• Decrease when zooming in
(Formula in paper)
Automatic Exploration
• Explore the regions in descending order of their uniqueness• k-NN anomaly detection step provides uniqueness ordering
Interactive Refinement
1. Locate similar undesired regions
2. Select a representative
3. Move the slider to adjust the coverage
4. Delete the selection
The spatial data structure indexes the regions and provides fast retrieval
After User Refinement
• The remaining 300 regions after 3 refinement interactions
Data Scalability
• GPU Out-of-core saliency computation• Break the image into tiles• Parallel Gaussian filtering on GPU• Filter overlapping boundary tiles to maintain continuity
• Saliency map storage• Fit and store ellipses of the salient regions• Do not store an extra image
• Tiled Image Viewer• View dependent mipmap image tiles loading and prefetching for
smooth pan and zoom
Royal Gorge Bridge (1.4 Gpix)
Cacti (4.0 GPix)
Mount Whitney (5.0 GPix)
Gigapan Community Tags
Grimsel Pass
Royal Gorge Bridge
Gigapan Community Tags
Cacti
Mount Whitney
Limitations
• Buffelgrass after fire• The “Original” cactus
• Tags with semantic information• Domain knowledge necessary
• Why are they tagged ?
PerformanceImage GigaPan Tags Detected
All Non Semantic
Grimsel Pass 35 32 25 (78%)
Royal Gorge 25 30 24 (80%)
Cacti 24 17 15 (88%)
Whitney 11 11 10 (90%)
• Each GPix takes 2.5 1 hours to preprocess(1 NVIDIA GeForce GTX 285 GPU and 1 CPU)
• Each interaction takes 10 ms
Conclusions
• First step on visual knowledge discovery on very large landscape images
• Visual Scalability: Sliding-Window Saliency• Information Scalability: Anomaly Detection• Data Scalability:Parallel filtering, Saliency Storage• Interactive Navigation
Future work
• There are a lot of very large images• Astronomy• Microscopy• Product inspection• Urban Scenes
1. Domain specific descriptors
2. Fast discovery of locally distinct regions.
3. Accurate Identification of globally unique regions.
Acknowledgements
• National Science Foundation: CCF 05-41120, CMMI 08-35572, CNS 09-59979
• NVIDIA CUDA Center of Excellence Program• Derek Juba, Sujal Bista, Rob Patro, Icaro da Cunha, Yang Yang,
Adil Yalcin, and the reviewers for improving this paper and presentation
• The Vis paper award committees
Thank you!
Questions ?
• Please see our websitesfor the paper and video:
• Cheuk Yiu Ip• www.cs.umd.edu/~ipcy/
• GVIL Research Highlights• www.cs.umd.edu/gvil/