chuchienshu / ultra-thin-prm

Weakly Supervised Instance Segmentation using Class Peak Response, in CVPR 2018 (Spotlight)

The reconstruction implementation of PRM by removing third-party dependency(i.e, Nest).

Motivation: An ultra-thin version of PRM, which aims at improving readability and expansibility.

Rule No.1: Never make code too complicated. 😂

Version info: pytorch 0.4.1, python 3.6

Training & Inference

Training:

python main.py --train True

Inference:

python main.py 

Sample result

Reference

@INPROCEEDINGS{Zhou2018PRM,
    author = {Zhou, Yanzhao and Zhu, Yi and Ye, Qixiang and Qiu, Qiang and Jiao, Jianbin},
    title = {Weakly Supervised Instance Segmentation using Class Peak Response},
    booktitle = {CVPR},
    year = {2018}
}

ultra-thin-prm's Issues

How to generate the json from the MCG proposals

Hi, thanks for your simplified implementation,
If I want to use this code on COCO, for the input data, how do you obtain the json segmentation files from the coarse MCG proposals, which further processing steps do you perform to obtain the binary map, and then transform it using rle_encode function?

I am looking forward to your reply,

Thanks,

.

最终还是放弃尝试了

How about the performance?

You are a hero!

Great job with this repo! It's so much simpler!

How to fine-tune?

Hi, thanks for this repository.
I want to fine-tune the model for a different number of object classes. How can I fine-tune only the head layers with my own small dataset?
Another question, how did you create the json files in the data folder?

FileNotFoundError: [Errno 2] No such file or directory: './snapshots/prm__0_checkpoint.pth.tar'

How can I get this file?

What does “**ctx.num_flags = 4**” mean in the ### prm.functions.peak_stimulation.py？

What does “ctx.num_flags = 4” mean in the ### prm.functions.peak_stimulation.py？

I want to test with pascal voc2012. How should I generate JSON files？

Could you offer the evaluation functions/files for reporting performance on val set?

How do I get proposal in .json format？

How to obtain "file_name.json" file of other images ?

In ./data/directory, there are many JSON files, but I don't know these files how to obtain. It seems that the code didn't explain this.

thank you for your answer