tao-train-mask-auto-label

Automate weakly-supervised segmentation by using a ViT-MAE backbone to generate segmentation masks from minimal point or box annotations, significantly reducing manual effort for training and inference.

Weakly-supervised segmentation is a technique that generates segmentation masks from minimal annotations like points or boxes. You need it when manual pixel-level labeling is too expensive for training deep learning models.

Yes, you need Docker and nvidia-container-toolkit to train, evaluate, and run inference on NVIDIA TAO MAL models, as the weakly-supervised segmentation workflow requires these containerized GPU environments.

Yes, you can use box annotations instead of point annotations. The weakly-supervised segmentation process accepts both minimal annotation types to generate accurate segmentation masks using the ViT-MAE backbone.

Run inference on NVIDIA TAO MAL models by executing the skill with required dataset paths in a Docker environment configured with nvidia-container-toolkit, utilizing the ViT-MAE backbone for efficient segmentation.

The ViT-MAE backbone approach for weakly-supervised segmentation requires Docker and nvidia-container-toolkit, meaning it is limited to environments supporting these dependencies and is optimized specifically for NVIDIA TAO tools.