Checklist

[gau-nernst]

Thoughts

• segmentation head (semantic and instance)
  • instance segmentation requires some kind of pooling?
• human pose head
• tracker.py -> (outside model). include BYTE algo. can work with just pure detector (centernet) or with appearance embeddings (FairMOT)
• utilities: inference image folder, video. method or cli script?
• heatmap loss
  • Gaussian: check other keypoint-based models e.g. HRNet
  • Generalized focal loss?
• bbox training samples: only GT center point (CenterNet); 3x3 center area (FCOS and YOLOX); Gaussian area with weighted sum (TTFNet). 5x5 center area?
• to cleanup: test folder, other datasets, config files
• update FairMOT. mechanism to handle classifier head
  • construction and how to inherit from CenterNet
  • if build from scratch, needs to separate detection decode functions from centernet
  • compute_loss: make a functions to get target boxes -> can reuse. but embeddings? which one will be used for training?
  • heatmap loss: compute every image?
• multi-head support. all detectors use multi-head. may improve performance significantly
  • probably don't separate objects by sizes to place on different feature maps. use some schemes like HRNet: train on all feature map levels, but during inference, fuse them together
• augmentations
  • torchvision
    • group by aspect ratio link. a good idea to avoid changing object's aspect ratios. does Albumentations have similar functions? does YOLOv5 have something in similar?
    • new recipe Improve the accuracy of Detection & Segmentation models by using SOTA recipes and primitives pytorch/vision#5307. still developing. presets and implementation
  • YOLOv5: recipe
    • mosaic
  • mmdet? link
  • TensorFlow? link
• model + train recipe with resnet50 should be competitive with other detectors

https://github.com/PaddlePaddle/PaddleDetection

TODO

Modelling

• Detections on multi-level feature maps

Training

• Refine transformation recipe for training and validation
• Implement Trivial Augment in Albumentations
• Implement Mosaic Augmentation

Done

• Change TrackEval dependency to the original github repo (https://github.com/JonathonLuiten/TrackEval)
• Use vision-backbones repo to replace current backbone + neck implementation (https://github.com/gau-nernst/vision-backbones)
• Replace custom config with Lightning CLI (which uses jsonargparse)
• Dataset: only support COCO to simplify pipeline
• Add DDP support