1. Sketch Tokens: A Learned Mid-level Representation for Contour and Object Detection CVPR2013 POSTER

2. Outline Introduction Method Results Discussion

3. Introduction Figure 1. Examples of sketch tokens learned from hand drawn sketches represented using their mean contour structure. Notice the variety and richness of the sketch tokens.

4. We typically utilize a few hundred tokens, which captures a majority of the commonly occurring edge structures. an efficient approach that can compute per-pixel token labelings in about one second per image. We propose a novel approach to both learning and detecting local edge-based mid-level features. Introduction

5. Defining sketch token classes Detecting sketch tokens Method

6. These include straight lines, t-junctions, y-junctions, corners, curves, parallel lines, etc. Defining sketch token classes

7. 1. Feature extraction 2. Classification Detecting sketch tokens

8. Two types of features are then employed: 1. features directly indexing into the channels 2. self-similarity features Feature extraction

10. 2. self-similarity features ：  The self-similarity features capture the portions of an image patch that contain similar textures based on color or gradient information.

13. Two considerations must be taken into account when choosing a classifier for labeling sketch tokens in image patches. First, every pixel in the image must be labeled, so the classifier must be efficient. Second, the number of potential classes for each patch ranges in the hundreds. Classification

16. 1. Contour detection 2. Object detection Results

17. Contour detection

18. We test our contour detector on the popular Berkeley Segmentation Dataset and Benchmark (BSDS500). Contour detection results

20. 1. INRIA pedestrian 2. PASCAL VOC 2007 Object detection

21. 1. INRIA pedestrian ： For pedestrian detection we use an improved implementation of Doll´ar et al. that utilizes multiple image channels as features for a boosted detector.

24. 2. PASCAL VOC 2007 ： Our final set of results use the PASCAL VOC 2007 dataset. The dataset contains real world images with 20 labeled object categories such as people, dogs, chairs, etc.

29. Discovering new sets of observable mid-level information that may be used for feature learning is an interesting and open question. We’ve explored several in this paper, but other tasks may also benefit. Discussion