facebookresearch/detectron2 v0.1.2

on GitHub

The pre-built wheels for this version have to be used with an official binary release of PyTorch 1.5.

Improvements:

• Add semantic segmentation models to PointRend
• Add examples to load a detectron2 model in C++
• New models that reproduce Rethinking ImageNet Pre-training
• Lots of new features in DensePose, see below
• Fix a few bugs in rotated box computation

Incompatible changes:

• When loading a checkpoint with resume_or_load(), training states like optimizer, start_iter will only be loaded when resume is True and the last checkpoint is found. This matches users’ expectations better
• .output_size in custom box head is renamed to .output_shape
• anchor_generator no longer duplicates the anchors for each image
• feature_strides and feature_channels attributes are removed from ROIHeads. Use the input argument input_shape instead.

New in DensePose:

• New evaluation metric (GPSm) that yields more reliable scores
• Panoptic FPN head implementation following Panoptic Feature Pyramid Networks
• DeepLabV3 head implementation following Rethinking Atrous Convolution for Semantic Image Segmentation
• Models with confidence estimation: implementation of the paper Correlated Uncertainty for Learning Dense Correspondences from Noisy Labels
• Contributions to the Model Zoo (for details please see the Model Zoo page):
  • Panoptic FPN and tuned hyperparameters for the original fully convolutional DensePose head: the performance of the existing R50 and R101 baselines improved by +10 AP
  • Panoptic FPN with DeepLabV3 head for DensePose: gives additional improvement of +2 AP
  • Models with confidence estimation
• Test time augmentations for DensePose (additional improvement of about +0.5 AP)