Download
DeepMask is an early, influential approach to class-agnostic object segmentation that learns to propose pixel-accurate masks directly from images. Instead of first generating boxes and then refining them, the network predicts a foreground mask and an “objectness” score for a given image patch, yielding high-quality segment proposals suitable for downstream detection or instance segmentation. The model is trained end-to-end to align mask shape with object extent, which markedly improves recall at a manageable number of proposals. In practice, DeepMask is run on an image pyramid with a sliding window, followed by non-maximum suppression to produce a compact set of candidates. A companion refinement model (SharpMask) sharpens the coarse predictions, recovering fine boundaries like thin limbs or object edges. The repository (in the original Torch/Lua stack) includes pretrained weights, training scripts, and evaluation utilities.
Features
- Class-agnostic mask proposal network that predicts both mask and objectness
- Multi-scale sliding-window inference with non-maximum suppression
- Optional refinement stage (SharpMask) for crisp, boundary-aware masks
- Torch/Lua implementation with pretrained models and scripts
- High recall with relatively few proposals for efficient downstream use
- Utilities to export proposals and integrate with detection pipelines
Project Samples
