DensePose CSE and DensePose Evolution

• DensePose Evolution pipeline, a framework to bootstrap DensePose on unlabeled data
  • InferenceBasedLoader with data samplers to use inference results from one model to train another model (bootstrap);
  • VideoKeyframeDataset to efficiently load images from video keyframes;
  • Category maps and filters to combine annotations from different categories and train in a class-agnostic manner;
  • Pretrained models for DensePose estimation on chimpanzees;
  • DensePose head training from partial data (segmentation only);
  • DensePose models with mask confidence estimation;
  • DensePose Chimps dataset for IUV evaluation
• DensePose Continuous Surface Embeddings, a framework to extend DensePose to various categories using 3D models
  • Hard embedding and soft embedding losses to train universal positional embeddings;
  • [Embedder](../(densepose/modeling/cse/embedder.py) to handle mesh vertex embeddings;
  • Storage for evaluation with high volumes of data;
  • Pretrained models for DensePose CSE estimation on humans and animals;
  • DensePose Chimps and DensePose LVIS datasets for CSE finetuning and evaluation;
  • Vertex and texture mapping visualizers;
• Refactoring of all major components: losses, predictors, model outputs, model results, visualizers;
  • Dedicated structures for chart outputs, chart outputs with confidences, chart results, CSE outputs;
  • Dedicated predictors for chart-based estimation, confidence estimation and CSE estimation;
  • Generic handling of various conversions (e.g. from outputs to results);
  • Better organization of various losses;
  • Segregation of loss data accumulators for IUV setting and CSE setting;
  • Splitting visualizers into separate modules;
• HRNet and HRFPN backbones;
• PoseTrack dataset;
• IUV texture visualizer