TESN: Transformers Enhanced Segmentation Network
Introduction
We proposed Transformers Enhanced Segmentation Network (TESN) which made the following improvements to the original Mask R-CNN to improve the accuracy of instance segmentation: a) TESN adopt multi-stage architecture to provide high-quality target detection. b) TESN Introduce Transformer into the mask head to provide high-quality segmentation.
The framework of the mask head. (a) mask head of original Mask R-CNN. (b) mask head of TESN. (c) convolution and up-sampling. (d) Transformer layer.
Results on MS COCO dataset:
Prepared to train
Environment
python 3.7.10 cuda 10.1 torch 1.5.0 mmcv-full 1.3.8 mmdet 2.18.0
File
1)Place fcn_mask_transformer_head.py and transformer.py under ../mmdetection/mmdet/models/roi_heads/mask_heads 2)Place TESN_COCO.py under ../mmdetection 3)add ‘from .fcn_mask_transformer_head import FCNMaskTransformerHead’ into ../mmdetection/mmdet/models/roi_heads/mask_heads/init.py
Train
Run ‘python tools/train.py TESN_COCO.py’
Test
Run ‘python tools/test.py TESN_COCO.py’
Acknowledgement
With greatly appreciation for open-mmlab for providing mmdetection source code.
