descriptron-vision

Mask and landmark prediction for Descriptron: torchvision detectors, SAM2-PAL propagation and DINOv3 landmark transfer.

pip install descriptron-vision
descriptron-train --task masks --coco-json annotations.json --img-dir images/ \
                  --output-dir out/ --total-iters 20000
descriptron-predict --checkpoint out/model_final_*.pth --coco-json annotations.json \
                    --img-dir images/ --annotations_out predictions.json

No compiler required. torch and torchvision are ordinary wheels on Linux, macOS and Windows. For CUDA, install torch from PyTorch's own index first:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install descriptron-vision

On a Mac the detectors can use the Apple GPU through PyTorch's MPS device, which is why this backend exists: Detectron2 has no macOS wheels and no Apple Silicon support.

Two optional dependencies that cannot be declared

PyPI rejects a git dependency in package metadata, so neither SAM2 nor Detectron2 is listed. Both are imported lazily, and the error names the command to install them. The Docker image carries both ready to run.

Choosing a backbone

--arch v1 (default) is maskrcnn_resnet50_fpn; --arch v2 is maskrcnn_resnet50_fpn_v2. v2 scores higher on COCO, but on a few hundred training images v1 won on 3 of 3 folds of a grouped species hold-out — its lighter two-FC box head converges further in the same budget. Measured on the data, not assumed from the model zoo.