v0.1.2
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 likeoptimizer
,start_iter
will only be loaded whenresume
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
andfeature_channels
attributes are removed fromROIHeads
. Use the input argumentinput_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)