Orals

Presentations that caught my eye in the 6 orals sessions from Tuesday to Thursday.

06/18/2019 (Tuesday)

Few shot learning

• Support set with few labeled sample per class, query set

new CNN architecture

• Kerverlutional Neural Network
• Polynomial kernel, gaussian kernel
• Why does ReLU get high confidence (overconfident) far away from the training data
• Neural rejuvenation: force dead neuron to reinitialize
• deep metric learning: contrastive vs triplet learning
• Classification –> dense image prediction
• AUtO-DeepLab: similar to auto FPN
• Learning loss for active learning Note
• Striking the Right Balance with Uncertainty

Scenes and Representation

domain adaptation

• d-SNE: generalize model from source domain to target domain
• ADVENT: Adversarial Entropy Minimization for Domain Adaptation in Semantic Segmentation
• Note Large-Scale Long-Tailed Recognition in an Open World
  • Dynamic Few-Shot Visual Learning without Forgetting (CVPR2018)
• Towards Scene Understanding: Unsupervised Monocular Depth Estimation With Semantic-Aware Representation:
  • train depth and semantic segmentation to help each other
  • train from different datasets
• Does CNN which perform well on ImageNet generalize well?

06/19/2019 (Wednesday)

Motion and tracking

• Use siamese network for tracking
• Note Deeper and Wider Siamese Networks for Real-Time Visual Tracking
  • ResNet causes performance drop. Why?
  • SiamDW: redesign residual blocks for tracking
• Modeling games with GNN (with inherent insensitivity to permutation)
• Detection and pose tracking
• SiamFC: Fully-Convolutional Siamese Networks for Object Tracking (ECCV 2016)
• Graph Convolutional Tracking

Recognition

• Notes Mask Scoring R-CNN –> can be adapted to object detection
  • mask Iou is low, but mask score is high
  • regress iou head for predicting mask iou
  • this improves high quality segmenation
• Adaptive NMS: Refining Pedestrian Detection in a Crowd
• Locating Objects Without Bounding Boxes
• Few-shot learning with localization in realistic settings Note
  • three parameter free techniques
  • Batch folding Note

06/20/2019 (Thursday)

Video Segmentation Note

• Improving Semantic Segmentation via Video Propagation and Label Relaxation
  • relax boundary in the gt label
• Accel: A Corrective Fusion Network for Efficient Semantic Segmentation on Video
  • 
  • 
  • 
  • Deep feature flow from MSRA
• Relation-Shape Convolutional Neural Network for Point Cloud Analysis
  • Seems like pointnet?
• BubbleNets: Learning to Select the Guidance Frame in Video Object Segmentation by Deep Sorting Frames
  • This is similar to an actor-critic model?
  • DAVIS2017 (first frame annotaiton)
• Collaborative Global-Local Networks for Memory-Efficient Segmentation of Ultra-High Resolution Images
  • can it be used for accelerating existing methods?
• Efficient Parameter-Free Clustering Using First Neighbor Relations Note
• A Generative Appearance Model for End-To-End Video Object Segmentation

Deep learning

• ABN (attention branch network)

Notes

• Main contribution template:
  • Proposed the problem
  • Created a method to address the problem
  • Contributed frist benchmark
• Video segmentation
  • full segmentation on key frames and label propagation for frames in-between
• FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search
  • latency LUT works really great!