• Co-author of this work.
• Formulation of zero-shot part discovery on a large-scale fine-grained shape segmentation benchmark.
• A learning-based agglomerative clustering framework that learns to do part proposal and grouping from training categories and generalizes to unseen novel categories

• Re-interpretation and connection of two top popular techniques proposed in CVPR 2018 (Non-local neural networks and SE-Net).
• Unification of these the two techniques into a general framework.
• Better instantiation of the general framework, which is about 50x faster than the non-local neural block, while achieving better accuracy than both techniques (non-local and SE-Net) on several recognition tasks such as ImageNet classification, COCO object detection and Kinetics action recognition.

• The first coherent and end-to-end framework for similarity measure which combines all of the appearance, motion and interaction cues.
• Properly redesign of feature representation for the tracklet-object pair.
• Achieve the state-of-the-art multi-object tracking (MOT) results on all of the MOT15-17 leaderboards using few bells and whistles.

This paper proposes the first end-to-end deep framework for high dynamic range (HDR) imaging of dynamic scenes with large-scale foreground motions.