Note: the email associated with the paper is no longer valid. Please contact me at yanfengliux@gmail.com
An extension of this work into multi-class multi-object tracking can be found in my thesis
Fig 1. Image and training ground truth: front class mask, occlusion class mask, front instance mask, occlusion instance mask
Fig 2. Network architecture
Fig 3. Failure case where there is a three-stack
Fig 4. Incomplete instance mask for the object at the bottom in the failure case
Fig 5. Mask R-CNN fails when there many objects of the same class and the bounding boxes are too crowded