Light-Weight RefineNet (in PyTorch) (Unofficial/Modified)
This repository provides unofficial/modified code for LightWeight-RefineNet from the paper Light-Weight RefineNet for Real-Time Semantic Segmentation, available here
Light-Weight RefineNet for Real-Time Semantic Segmentation
Vladimir Nekrasov, Chunhua Shen, Ian Reid
In BMVC 2018
Official code can be picked from the author's implementation available here
Changes w.r.t. the official code
I've modified the code to work on CityScapes dataset, and also visualize the predicted labels. Using the training configuration in src/config.py, I could obtain the best valdiation score of 75.28% on this dataset. For near replication, you can either -
- Train the model yourself. Simply run
source train.sh - Download the pre-trained checkpoint from here and place it at
ckpt/run_20190219/. Then, simply runsource test.sh. It will produce the aforementioned score and also save the predicted labels atoutputs/run_20190219/cityscapes/
For either, ensure that you've downloaded the dataset first (from here) and modified the directory params accordingly (in src/config.py or src/test.py depending upon whether your are training or testing). Feel free to experiment with other training configurations, and kindly inform me if you can achieve a better validation/test score.