reagan1311 / asgnet Goto Github PK

Adaptive Prototype Learning and Allocation for Few-Shot Segmentation (CVPR 2021)

Home Page: https://reagan1311.github.io/asgnet/

Python 99.09% Shell 0.91%

few-shot-segmentation prototypical-learning pascal-5i coco-20i cvpr2021

asgnet's Introduction

ASGNet

The code is for the paper "Adaptive Prototype Learning and Allocation for Few-Shot Segmentation" (accepted to CVPR 2021) [arxiv]

Overview

data/ includes config files and train/validation list files
model/ includes related model and module
tool/ includes training and testing scripts
util/ includes data processing, seed initialization

Usage

Requirements

python==3.7, torch==1.6, scipy, opencv-python, tensorboardX

Dataset

Prepare related datasets: Pascal-5ⁱ (VOC 2012, SBD) and COCO-20ⁱ (COCO 2014)

Pre-trained models

Pre-trained backbones and models can be found in Google Driver
Download backbones and put the pth files under initmodel/ folder

Test and Train

Specify the path of datasets and pre-trained models in the data/config file

Use the following command

sh tool/test.sh|train.sh {data} {model} {split_backbone}

E.g. Test ASGNet with ResNet50 on the split 0 of PASCAL-5i:

sh tool/test.sh pascal asgnet split0_resnet50

Citation

Please consider citing the paper if you find it useful:

@inproceedings{li2021AdaptivePL,
  title={Adaptive Prototype Learning and Allocation for Few-Shot Segmentation},
  author={Gen Li and Varun Jampani and Laura Sevilla-Lara and Deqing Sun and Jonghyun Kim and Joongkyu Kim},
  booktitle={CVPR},
  year={2021}
}

References

The code is based on semseg and PFENet. Thanks for their great work!

asgnet's People

Contributors

Stargazers

Watchers

Forkers

lemoshu zots0127 rituraj-commits trendingtechnology mess0223 paulinoaranda wingkeungm glbreeze holmes-gu glenggg vineet-the-git donglongzi sasenczy wasongiwi

asgnet's Issues

datasets

Dear author：
When I use the Pascal-5i dataset，label image is colorful，after label_class = np.unique(label).tolist()，it can't be satisfied in the sub_list

TypeError: 'int' object is not iterable

Hello, thanks for work!
I have some problem when I try to train the model with pascal set. I follow the step show in the github page, and run:
sh tool/train.sh pascal asgnet split0_resnet50
but I get the error as follow:

[2022-02-22 16:10:06,166 INFO train.py line 329 19996] >>>>>>>>>>>>>>>> Start Evaluation >>>>>>>>>>>>>>>> Traceback (most recent call last): File "/opt/conda/envs/dual_learning/lib/python3.8/runpy.py", line 194, in _run_module_as_main return _run_code(code, main_globals, None, File "/opt/conda/envs/dual_learning/lib/python3.8/runpy.py", line 87, in _run_code exec(code, run_globals) File "/home/tool/train.py", line 450, in <module> main() File "/home/tool/train.py", line 91, in main main_worker(args.train_gpu, args.ngpus_per_node, args) File "/home/tool/train.py", line 207, in main_worker loss_val, mIoU_val, mAcc_val, allAcc_val, class_miou = validate(val_loader, model, criterion) File "/home/tool/train.py", line 421, in validate allAcc = sum(intersection_meter.sum) / (sum(target_meter.sum) + 1e-10) TypeError: 'int' object is not iterable

I try to solve this, and I found the "intersection_meter.sum" belongs to "int" type, which leads to the error. As normal, the 'intersection_meter.sum' should be a list or array? Where ma'y I go wrong?
Thanks a lot1

model

Hello,I want to konw if the pretained models in Google Drive is the best.pth in training.
Because I ues it in test.py,loss is 0.66.

The mIoU of split0 is lower than that in the paper

Thank for sharing your code. When I run the training by
sh tools/train.sh pascal asgnet split0_resnet50
with your default setting and run the testing by
sh tools/test.sh pascal asgnet split0_resnet50
using the checkpoint final.pth, I get the miou of 0.5389 which is much lower than 58.84 in the paper. Do I miss some steps ?

how to prepare the datasets?

Hello, I'm sorry to disturb you, but I am a beginner and I have encountered some problems when preparing the COCO and PASCAL data sets.

After I downloaded the VOC2012 data set, I found that there were only about 2000 pictures, but there were more than 5000 data in ASGNet/data/lists/pascal/voc_sbd_merge_noduplicate.txt. And then, I think maybe I should use VOC2012Aug, but there are more than 10000 pictures in VOC2012Aug.
I downloaded the COCO data train/test/val in the warehouse link and decompressed them, but I did not find the data label.
I will be very happy if I can get your help, thank you.

where is coco2012/SegmentationClassAug folder？

how to get the folder of coco2012/SegmentationClassAug?
I download the coco2012 dataset on the internet, but it only has SegmentationClass, without SegmentationClassAug.

View the result in the original image

When I want to view the prediction results of the pictures in the test set, I found that the code I wrote to draw the output results on the original pictures always has bugs and ambiguities. Can I get the code that compares the results in the paper?

When I am performing image inference, the input is the image to be detected and a picture of the target class and its mask picture, then how can I use a picture to quickly detect different pictures, what I actually want to say is: how It is better to detect unknown pictures, not that I can only use one picture for verification.

Hope to provide an idea

Visualization of Super-pixel Regions

Hi, thank you for providing such excellent work, can you supply the code to visualize the clustered regions like this sample:

The 5-shot evaluation is problematic

Hi, thanks for your work.

I have reproduced the 1-shot results that are close to the ones in your paper.

However, when I directly test the model trained with 1-shot in the 5-shot evaluation setting, the result is far from yours.

For example, on fold 0, the 1-shot result is 59.2 and 5-shot I only got 43.73.

Could you help fix this issue?

Thankyou.

Project request

Hello, first of all, I would like to express my compliments to this network structure you proposed. I think your paper has given me a lot of help, because I am still a beginner and cannot fully implement the project, so I hope you can give me a complete project, I want to see the complete running effect of this project, and hope to get your reply

Category settings in training and testing

Hello thanks for making the code available.

I have a question. Why are the classes in the configuration file set to 2? These are two specific categories? Or does it have a special meaning?