DensePose CSE and DensePose Evolution
- DensePose Evolution pipeline, a framework to bootstrap
DensePose on unlabeled data
InferenceBasedLoaderwith data samplers to use inference results from one model to train another model (bootstrap);VideoKeyframeDatasetto 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 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