Mask R-CNN in TensorFlow, assembled but not yet road-tested

What it does CharlesShang/FastMaskRCNN is an attempt to reproduce the Mask R-CNN paper in TensorFlow 1.x, stitching together ResNet-50, a Feature Pyramid Network, ROIAlign, and a mask head. It includes COCO data ingestion, a Cython-accelerated anchor layer, and a training script, but stops short of claiming a working end-to-end pipeline.

The interesting bit The README’s candor is refreshing: the author explicitly warns that “there are certainly some bugs” and labels testing and debugging as still “in progress.” Most architectural boxes are checked—ROIAlign, FPN, mask layer—yet the actual COCO training and evaluation line item remains stubbornly empty.

Key highlights

• Reproduces the full Mask R-CNN architecture in TensorFlow (>= 1.0.0)
• Implements ROIAlign, FPN, and the mask head (with heavy borrowing from TFFRCNN and py-faster-rcnn)
• Cython-accelerated anchor layer for a modest speed bump
• Explicit bug disclaimer and open call for contributions
• TODO list shows most modules combined, but end-to-end training is not yet verified

Caveats

• Training and evaluation on COCO is still unchecked on the TODO list
• README warns there are “certainly some bugs” and testing is ongoing
• Cannot yet train on more than two images at once, per the TODO

Verdict Worth a look if you want to study how Mask R-CNN pieces fit together in legacy TensorFlow, but anyone needing a reliable, ready-to-train implementation should probably look elsewhere—at least until the COCO training checkbox gets filled.