TODO

Backbone

• Hourglass [Paper, ECCV 2016]
  • Usage: default (--num-stack 1 --hourglass-inch 128 --increase-ch 0)
• VGG
• ResNet
• CSPDarknet53

Neck

• SPP (Spatial Pyramid Pooling) [Paper, TPAMI 2015]
  • Usage: --neck-pool SPP (default: --neck-pool None)
• SAM (Spatial Attention Module) [Paper, ECCV 2018]
• PAN (Path Aggregation Network) [Paper, CVPR 2018]

Head

• CenterNet [Paper, Arxiv 2019]
  • Usage: default
• YOLOv3 [Paper, Arxiv 2018]
• FSAF (Feature Selective Anchor-Free) [Paper, 2019 CVPR]

Layer

• Deformable Convolution

BoF (Bag of Freebies)

• CutMix data augmentation
• Mosaic data augmentation
• DropBlock regularization
• Cosine annealing scheduler [Paper, ICLR 2017]
• Random training shapes
  • Usage: --multiscale-flag --multiscale 320 640 64
• Focal Loss
  • Usage: default (--focal-alpha 2.0 --focal-beta 4.0)
• Smooth L1 Loss
• DIoU Loss
• IoU Loss
• Group Normalization

BoS (Bag of Specials)

• Mish [Paper, ECCV 2020]
  • Usage: --activation Mish(default: --activation ReLU)
• SPP (Spatial Pyramid Pooling) [Paper, TPAMI 2015]
  • Usage: --pool SPP (default: --pool Max)
• Soft-NMS [Paper, ICCV 2017]
  • Usage: --nms soft-nms (default: --nms nms)
• DIoU-NMS

Implementation

• normalized (relative) coordinate prediction
  • Usage: --normalized-coord (The loss weight needs to be modified when using this flag.)
• Resume training
• Single image prediction
• Load pretrained model
  • Load pretrained backbone
• Set learning rate per module differently
• Validation durning training
• Clean code
• Loading a torchscript model in c++
• Maintain aspect ratio of the input when evaluation
• Multi-view voting
  • Multisclae prediction
  • Horizontal Flip prediction
• Deployment with Ainize
• Multivariate Gaussian to generate heat map
• Upgrade to Pytorch 1.7.0