- SuperGlue: Learning Feature Matching With Graph Neural Networks
- Reinforced Feature Points: Optimizing Feature Detection and Description for a High-Level Task
- Cascaded Deep Monocular 3D Human Pose Estimation With Evolutionary Training Data
- Self-Supervised Deep Visual Odometry With Online Adaptation [Generalizes to unseen data]
- 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
- 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
