Parsing Semantic Parts of Cars Using Graphical Models and Segment Appearance Consistency
In Proceedings British Machine Vision Conference 2014
AbstractThis paper addresses the problem of semantic part parsing (segmentation) of cars, i.e. assigning every pixel within the car to one of the parts (e.g. body, window, lights, license plates and wheels). We formulate this as a landmark identification problem, where the set of landmarks specify the boundaries of the parts. A novel aspect of our model is that we dynamically couple the landmarks to a hierarchy of segments (obtained by Segmentation by Weighted Aggregation). This enables the model to use the appearance of visual segments while parsing the car and, in particular, to enforce appearance consistency between segments within the same part. The model is learnt using latent SVM. Parsing the car is performed by dynamic programming, including finding the optimal coupling between landmarks and segments in the hierarchy. We evaluate our method on a new dataset, PASCAL VOC 2010 where we have hand-labelled the positions of the parts, and on the car subset of 3D Object Category dataset (CAR3D). We show good results and, in particular, quantify the effectiveness of using the segment appearance consistency in terms of accuracy of part localization and segmentation.
FilesExtended Abstract (PDF, 1 page, 2.0M)
Paper (PDF, 12 pages, 4.5M)