A Shape Context Based Car Detection with Hypothesis Pruning
Transkript
A Shape Context Based Car Detection with Hypothesis Pruning
Introduction Related Work Methodology Results A Shape Context Based Car Detection with Hypothesis Pruning Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç January 13, 2010 Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Introduction Related Work Methodology Overview Top-Down Recognition Codebook Building Improved Shape Context Hypothesis Generation Hypothesis Pruning False Positive Elimination Unification of Related Hypothesis Results Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Car detection and recognition is a key topic where many research areas; robotics, navigation, surveillance benefit from. Cars can vary greatly by their: I shape I color I size I tires I headlights ... In our study, a car detection methodology using shape context(SC) feature is adopted Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Existing methodologies for car detection can be categorized as knowledge-based, motion-based and stereo-based. Our implementation uses knowledge-based car detection. Here are several previous work on the subject: I Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in crowded scenes. In: CVPR. (2005) Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Existing methodologies for car detection can be categorized as knowledge-based, motion-based and stereo-based. Our implementation uses knowledge-based car detection. Here are several previous work on the subject: I Belongie, S., Malik, J., Puzicha, J.: Shape matching and object recognition using shape contexts. IEEE Trans. Pattern Anal. Mach. Intell. 24(4) (2002) Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning The methodology has three parts: I I Codebook Building Top-Down Recognition I I I Improved Shape Context Hypothesis Generation Hypothesis Verification Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning The methodology has three parts: I I Codebook Building Top-Down Recognition I I I Improved Shape Context Hypothesis Generation Hypothesis Verification Pruning Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning Codebook Building At this stage, using a set of images and their masks, a training is performed where each feature vector correspond to a codebook entry. For point pi , codebook entry cei is as follows; cei = (ui , δi , mi , wi ) I ui : shape context feature vector I δi : position w.r.t. object center I mi : binary mask for patch around pi I wi : weight mask of mi Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç (1) A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning Shape Context Shape Context feature describe shapes to allow shape similarity measuring and point correspondence matching. I Select n point on the edges. I For each pi of n points, create n − 1 vectors from pi to all remaining points. I Create a histogram using a simple binning scheme: hi (k) = #{q 6= pi : (q − pi ) ∈ bin(k)} Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç (2) A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning Improved Shape Context I Angular Blur: When dense bins are used, even similar images can differ in histograms. I Mask Function: To eliminate the noise due to background textures masking is introduced. Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning Hypothesis Generation Algorithm 1. Compare each shape context(SC) feature with every codebook entry to predict possible object center. 2. Accumulate matching scores over whole image. 3. Predict points with maximum scores as possible object centers. Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning False Positive Elimination I Need a way to prune false positives. I A threshold value is estimated emprically where it eliminates the hypothesis with low score values. I Threshold score value is estimated as 65. Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Overview Top-Down Recognition Hypothesis Pruning Unification of Related Hypothesis I Improve the hypotheses by unifying related ones 1. Calculate the overlapping area ratios between each pairs of hypotheses 2. Select the ones having a ratio higher than a threshold and combine them Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Dataset In order to have a clean training dataset, we have collected over 400 images from Yahoo! Autos. These photos are taken from 5 different view points, and by taking horizontal flips of related images, we ended up 8 different poses, a total of 720 images. Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Sample Results Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Sample Results Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results References L. Wang, J. Shi, G. Song, and I.-F. Shen: ”Object detection combining recognition and segmentation,” 2007, pp. 189-199. Belongie, S., Malik, J., Puzicha, J.: Shape matching and object recognition using shape contexts. IEEE Trans. Pattern Anal. Mach. Intell. 24(4) (2002) Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in crowded scenes. In: CVPR. (2005) Yahoo! Autos, http://autos.yahoo.com/ Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning Introduction Related Work Methodology Results Any Questions ? Işıl Doğa Yakut, Cansın Yıldız, Sefa Kılıç A Shape Context Based Car Detection with Hypothesis Pruning