Visual Understanding

Image Understanding

• Optimizing Object Detection

Video Understanding

Video Object Detection and Tracking

Datasets:

• ImageNet DET: object detection from images with classes including VID
• ImageNet VID: video object detection with instance IDs
• MOT16: 14 video sequences filmed in unconstrained environments for tracking

Code:

• Detect to Track and Track to Detect
• Object detection from video tubelets with convolutional neural networks

Action Recognition

Datasets:

• AVA: fine-grained audiovisual annotations of long human action videos
• ActivityNet: large-scale video benchmark for human activity understanding
• Kinetics: large-scale high-quality action videos from YouTube up to 700 classes
• THUMOS: action recognition in temporally untrimmed video
• Jester: humans performing hand gestures in front of a webcam
• Something-Something: humans performing basic actions with everyday objects
• Charade: people recording videos at home, acting out casual everyday activities
• Charade-Ego: indoors activities through AMT recorded from both third and first person

Object Detectors

Combo:

• mmdetection
  • SOTA model zoo including Faster/Mask R-CNN, RetinaNet, DCN-v2, etc. but YOLO
  • up-to-date optimizations including Group Normalization (GN) and Weight Standardization (WS)

One Stage:

• YOLOv3: fastest
• RetinaMask: one stage Mask R-CNN

Two Stage:

• 1.6K Class Detectron Training Protocol w/ Attributes
  • discussion
• Mask R-CNN
• Faster R-CNN

Compilations

• Awesome-Multi-Camera-Network
• List of Video Object Detections
• List of Temporal Action Detection
• List of Action Recognitions
• CVPR 2018 Person Re-identificationsted