This repository is for reproducing the results of the fallowing paper
S. R. Alvar and I. V. Bajić, "Multi-task learning with compressible features for Collaborative Intelligence".
accepted for presentation at IEEE ICIP'19, Taipei, Taiwan, Sep. 2019.
To run the code you need Pytorch = 0.4 and the Python=3.5.2. The model weights are in here. The name of the weight file is task_(encoder_type= 256 or 512)_(1=baseline or 2=proposed).pkl
There are 4 sets of experiments:
- Encoder (512x8x16) baseline
- Encoder (512x8x16) proposed
- Encoder (256x16x32) baseline
- Encoder (256x16x32) proposed
For each experiment you need to run the experiments for 5 times (PNG --quality 0 , JPEG --quality:95,90,85,80). The byte number for the 500 sequences and the average performance will be printed. (./results includes the printed result and the excel file is the summary of 20 txt files and has the curves plotted). To see the results of each experiment use a command similar to the fallowing
python ./write_features_3_tasks_512.py --dataset cityscapes --img_rows 256 --img_cols 512 --resume_feature ./feature_model_512_1.pkl --resume_segment ./segment_model_512_1.pkl --resume_reconstruct ./reconstruct_model_512_1.pkl --resume_depth ./depth_model_512_1.pkl --quality 0
For encoder 256x16x32 change 512 to 256.
Cityscapes dataset is needed and you should update the directory of the dataset in the config.json file.
After the results obtained, you can use the bjontegaard2.m to get the results in table 1.
References:
Parts of the code is based on the codes from:
https://github.com/meetshah1995/pytorch-semseg
https://www.mathworks.com/matlabcentral/fileexchange/41749-bjontegaard-metric-calculation-bd-psnr