SALPROP: SALIENT OBJECT PROPOSALS VIA AGGREGATED EDGE CUES
1 like132 views
1) The document proposes a novel object proposal generation scheme using a graph-based salient edge classification framework that utilizes edge context. 2) It uses fewer bounding boxes than other methods to accurately capture prominent objects while maintaining the saliency order of proposals. 3) Experimental results show the proposed SalProp technique achieves over 25% recall with only 1 proposal at IOU=0.5, outperforms other state-of-the-art methods, and has computational speedups.
![RESEARCH POSTER PRESENTATION DESIGN © 2012
www.PosterPresentations.com
We propose a novel object proposal generation scheme by
formulating a graph-based salient edge classification framework
that utilizes the edge context.
Key Features:
• Use fewer number of bounding boxes for good coverage of the
prominent objects contained in the image.
• Maintain order of saliency in the object proposals
Problem Statement
INTRODUCTION
Prerana Mukherjee*1, Brejesh Lall1, Sarvaswa Tandon2
*e-mail: eez138300@ee.iitd.ac.in
SALPROP: SALIENT OBJECT PROPOSALS VIA AGGREGATED EDGE CUES
1Department of Electrical Engineering, Indian Institute of Technology, Delhi, India; 2National Institute of Technology Goa, India
at IEEE International Conference on Image Processing (ICIP 2017), Beijing China
RESULTS PROPOSED APPROACH
CONCLUSION
• Novel object proposal generation algorithm operating in a
computationally efficient learning based setting where the
salient object edge density inside the bounding box is
analyzed to score the proposal set.
• High recall rates with lesser number of proposals with
varying IoU thresholds and subsequently making it more
reliable in context of competing methods.
• Ranked the key objects according to their saliency
REFERENCES
Fig. 1. The SalProp Framework. Given any RGB image, we
generate proposals ranked in the order of saliency. Green boxes
contain the most salient objects having higher rank and blue
boxes contain less salient objects and are ranked lower in the
proposal set. The number assigned to each box indicates its
saliency ranking in the proposal pool.
Input Image
Edge
Detection
(Oriented
Edge Forests)
Edge Saliency
Computation
(Bayesian
Framework)
Object/Non-Object
Edge Classification
(Conditional
Random Fields)
Window Scoring
SalProp Framework
Fig. 2. (a) Original image (b) Edge map using OEF (c) After
NMS and thresholding (d) Bayesian Probabilistic edge map
(indicating saliency of edge segments)
• SalProp is the best technique at lower number of proposals achieving over 25% and 19%
recall with only 1 window at IoU=0.5 and 0.6 respectively.
• At IoU=0:7, SalProp outperforms Rahtu [5] by 3.46%, Selective Search [4] by 5.16%,
Objectness [3] by 7.32%, Randomized Prim’s [6] by 8.71%, GOP [17] by 22.36%, Rigor [11]
by 23.46%, Rantalankila [7] by 30.05% and Perceptual Edge [9] by 30.35% at top-10
proposals.
• Outperforms objectness by 2%, 6% and 30% at IoU thresholds 0.5, 0.6 and 0.7 respectively.
• Comparable performance to EdgeBoxes while having a computational speedup of 5x over
MCG.
Fig. 4. Top Row: SalProp, Bottom Row: EdgeBoxes.
Closest bounding boxes (blue) having maximum overlap
with the ground truth boxes (green).
Fig. 3. (a) NMS cut-off threshold for highest recall value at varying IoU on validation set images. (b)-(d) The detection rate vs. the
number of bounding box proposals for varying IoU = 0:5, 0:6 and 0:7 on test set images.
[1] C. L. Zitnick, and Piotr Dollár. "Edge boxes: Locating object
proposals from edges." In European Conference on Computer Vision, pp.
391-405. Springer, Cham, 2014.
[2] Y. Qi, Yi-Zhe Song, Tao Xiang, Honggang Zhang, Timothy
Hospedales, Yi Li, and Jun Guo. "Making better use of edges via
perceptual grouping." In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 1856-1865. 2015.
[3] J. Pont-Tuset, Pablo Arbelaez, Jonathan T. Barron, Ferran Marques,
and Jitendra Malik. "Multiscale combinatorial grouping for image
segmentation and object proposal generation." IEEE transactions on
pattern analysis and machine intelligence 39, no. 1 (2017): 128-140.
• Object localization with high degree of precision is a
challenging task
• It is usually solved by
– Using feature statistics
– Generic object region proposals
– Deep learning (requires too much training data)
– Exploit Edges
• Edges capture most of the shape information thus preserving
important structural properties contained in the image.](https://image.slidesharecdn.com/icip2017-180306091918/85/SALPROP-SALIENT-OBJECT-PROPOSALS-VIA-AGGREGATED-EDGE-CUES-1-320.jpg)
SALPROP: SALIENT OBJECT PROPOSALS VIA AGGREGATED EDGE CUES
- 1. RESEARCH POSTER PRESENTATION DESIGN © 2012 www.PosterPresentations.com We propose a novel object proposal generation scheme by formulating a graph-based salient edge classification framework that utilizes the edge context. Key Features: • Use fewer number of bounding boxes for good coverage of the prominent objects contained in the image. • Maintain order of saliency in the object proposals Problem Statement INTRODUCTION Prerana Mukherjee*1, Brejesh Lall1, Sarvaswa Tandon2 *e-mail: eez138300@ee.iitd.ac.in SALPROP: SALIENT OBJECT PROPOSALS VIA AGGREGATED EDGE CUES 1Department of Electrical Engineering, Indian Institute of Technology, Delhi, India; 2National Institute of Technology Goa, India at IEEE International Conference on Image Processing (ICIP 2017), Beijing China RESULTS PROPOSED APPROACH CONCLUSION • Novel object proposal generation algorithm operating in a computationally efficient learning based setting where the salient object edge density inside the bounding box is analyzed to score the proposal set. • High recall rates with lesser number of proposals with varying IoU thresholds and subsequently making it more reliable in context of competing methods. • Ranked the key objects according to their saliency REFERENCES Fig. 1. The SalProp Framework. Given any RGB image, we generate proposals ranked in the order of saliency. Green boxes contain the most salient objects having higher rank and blue boxes contain less salient objects and are ranked lower in the proposal set. The number assigned to each box indicates its saliency ranking in the proposal pool. Input Image Edge Detection (Oriented Edge Forests) Edge Saliency Computation (Bayesian Framework) Object/Non-Object Edge Classification (Conditional Random Fields) Window Scoring SalProp Framework Fig. 2. (a) Original image (b) Edge map using OEF (c) After NMS and thresholding (d) Bayesian Probabilistic edge map (indicating saliency of edge segments) • SalProp is the best technique at lower number of proposals achieving over 25% and 19% recall with only 1 window at IoU=0.5 and 0.6 respectively. • At IoU=0:7, SalProp outperforms Rahtu [5] by 3.46%, Selective Search [4] by 5.16%, Objectness [3] by 7.32%, Randomized Prim’s [6] by 8.71%, GOP [17] by 22.36%, Rigor [11] by 23.46%, Rantalankila [7] by 30.05% and Perceptual Edge [9] by 30.35% at top-10 proposals. • Outperforms objectness by 2%, 6% and 30% at IoU thresholds 0.5, 0.6 and 0.7 respectively. • Comparable performance to EdgeBoxes while having a computational speedup of 5x over MCG. Fig. 4. Top Row: SalProp, Bottom Row: EdgeBoxes. Closest bounding boxes (blue) having maximum overlap with the ground truth boxes (green). Fig. 3. (a) NMS cut-off threshold for highest recall value at varying IoU on validation set images. (b)-(d) The detection rate vs. the number of bounding box proposals for varying IoU = 0:5, 0:6 and 0:7 on test set images. [1] C. L. Zitnick, and Piotr Dollár. "Edge boxes: Locating object proposals from edges." In European Conference on Computer Vision, pp. 391-405. Springer, Cham, 2014. [2] Y. Qi, Yi-Zhe Song, Tao Xiang, Honggang Zhang, Timothy Hospedales, Yi Li, and Jun Guo. "Making better use of edges via perceptual grouping." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1856-1865. 2015. [3] J. Pont-Tuset, Pablo Arbelaez, Jonathan T. Barron, Ferran Marques, and Jitendra Malik. "Multiscale combinatorial grouping for image segmentation and object proposal generation." IEEE transactions on pattern analysis and machine intelligence 39, no. 1 (2017): 128-140. • Object localization with high degree of precision is a challenging task • It is usually solved by – Using feature statistics – Generic object region proposals – Deep learning (requires too much training data) – Exploit Edges • Edges capture most of the shape information thus preserving important structural properties contained in the image.