[Project Page] [Paper] [Poster]

This is an update on GlassSemNet from which the model is now able to produce semantic segmentation predictions for visualization purpose.

Segmentations produced by the semantic backbone illustrated that the model was able to recognize object semantics. Features (layers 2 and 4) from intermediate layers were extracted and fed to auxiliary classifiers.

Some minor improvements were obtained after changing the semantic backbone from ResNet50 to ResNext50_32x4d.

Inference and visualization scripts with respective required input directories. Trained model weights (v2) available for inference and testing.

# To view the segmentation results, pass the 'SEMANTIC' flag argument into the scripts

> python predict.py -c CHECKPOINT -i IMAGE -o OUTPUT [-s SEMANTIC]

> python visualize.py -i IMAGE -p PREDICTION -o OUTPUT [-s SEMANTIC] 

Evaluation script for performance assessment. Results by GlassSemNetv2 available for reference.

> python evaluation.py -p PREDICTION -gt GROUNDTRUTH

@article{neurips2022:gsds2022,
  author    = {Lin, Jiaying and Yeung, Yuen-Hei and Lau, Rynson W.H.},
  title     = {Exploiting Semantic Relations for Glass Surface Detection},
  journal   = {NeurIPS},
  year      = {2022},
}