RefineMask

RefineMask + DCL

We apply DCL by modifying the official code of RefineMask which is based on mmdetection instead of detectron2. Since few changes are required, we only release the modified parts.

One can apply DCL to other models of the mmdetection library with similar modifications.

Installation

• First, install RefineMask following its official repository. Note that our code was tested on commit 633ed2b.

• Replace the corresponding files (i.e. mmdet/models/roi_heads/refine_roi_head.py etc).

Training

Run the following commands to train RefineMask + DCL on Cityscapes dataset. Note that we use 1.4 as the loss weight for DCL, which generates slightly better results than the paper (1.0).

./scripts/dist_train.sh \
    configs/refinemask/cityscapes/r50-refinemask-dcl-w14.py\
    8 \
    work_dirs/r50-refinemask-dcl-w14-cityscapes

Evaluation

• Generate Cityscapes-style results on validation set:

  ./scripts/dist_test.sh \
      configs/refinemask/cityscapes/r50-refinemask-dcl-w14.py \
      r50-refinemask-dcl-w14-cityscapes.pth \
      8  \
      --format-only \
      --options "txtfile_prefix=./r50-refinemask-dcl-w14-val"

• Evaluate the results:

  CITYSCAPES_DATASET=data/cityscapes/ \
  CITYSCAPES_RESULTS=r50-refinemask-dcl-w14-val \
  python -m cityscapesscripts.evaluation.evalInstanceLevelSemanticLabeling