Orals

ORAL 2.1A: 3D FROM MULTIVIEW AND SENSORS (3)

• SuperGlue: Learning Feature Matching With Graph Neural Networks
• Reinforced Feature Points: Optimizing Feature Detection and Description for a High-Level Task

Oral 2.2A: Face, Gesture, and Body Pose (2)

• Cascaded Deep Monocular 3D Human Pose Estimation With Evolutionary Training Data
• Self-Supervised Deep Visual Odometry With Online Adaptation [Generalizes to unseen data]

Oral 2.2C: Representation Learning

• Circle Loss: A Unified Perspective of Pair Similarity Optimization

Oral 2.2B: Motion and Tracking (1)

• GNN3DMOT: Graph Neural Network for 3D Multi-Object Tracking With 2D-3D Multi-Feature Learning

Poster session

• HigherHRNet: Scale-Aware Representation Learning for Bottom-Up Human Pose Estimation
• A Multi-Task Mean Teacher for Semi-Supervised Shadow Detection [Shadow segmentation]
• The Devil Is in the Details: Delving Into Unbiased Data Processing for Human Pose Estimation [data processing miss by 1?]
• Siamese Box Adaptive Network for Visual Tracking
• How to Train Your Deep Multi-Object Tracker
• [End-to-End Camera Calibration for Broadcast Videos]
  • calibration for sports, to find basketball and football courts
• Physically Realizable Adversarial Examples for LiDAR Object Detection
  • Uber ATG
  • Placing an object on top of vehicle roof leads to misdetection
  • Adversarial objects on rooftop are not uncommon for lidar object detection.
• [Cylindrical Convolutional Networks for Joint Object Detection and Viewpoint Estimation] ()
• Fast(er) Reconstruction of Shredded Text Documents via Self-Supervised Deep Asymmetric Metric Learning
  • reconstruction of shredded documents
• Learning to Evaluate Perception Models Using Planner-Centric Metrics
  • Not all cars are equally important
  • Generate planning based on perception results and evaluate trajectory from GT.
• DeFeat-Net: General Monocular Depth via Simultaneous Unsupervised Representation Learning
  • more robust depth estimation under diff lighting conditions
• SpeedNet: Learning the Speediness in Videos
  • Only classify normal speed or speed up
  • adaptive speedup of video: looks more natural, less jittering
  • self-supervised learning for video understanding
  • video retrival: similar motion pattern
  • spatial temporal visualization
  • This seems to be quite similar to Video Playback Rate Perception for Self-Supervised Spatio-Temporal Representation Learning
• Learning to Measure the Static Friction Coefficient in Cloth Contact
  • Predicting friction parameter of fabrics with video
  • simulator to generate synthetic datasets
  • conv + LSTM + fc
• 15 Keypoints Is All You Need
  • We track human poses with transformers that are input keypoint sequences. This achieves SOTA accuracy while using 500x fewer FLOPS than optical flow.
• Joint Spatial-Temporal Optimization for Stereo 3D Object Tracking:
  • deep learning + temporal consistency for optimization
• Warping Residual Based Image Stitching for Large Parallax
  • Parallax robust image stithing
  • Review of image stiching on Zhihu