lannguyen0910/deep-efficient-person-reid

Experiment about Deep Person Re-identification with EfficientNet-v2

👋 Hi, I’m Hoang-Lan. I’m a Master's student in Germany, formerly an apprentice Data Engineer.

🔭 I’m interested in data engineering, cloud computing, machine learning, and deep learning.

📫 My email: [email protected].

🙋‍♂️ More about me: ✨ https://lannguyende.com/ ✨

👨‍💻 Tech stacks

👨‍🏫 Profile 3D

☰ Read More

🏆 Trophies

📈 My Github Stats

---

⚔️ Activities

---

📞 Contact me at

deep-efficient-person-reid's Issues

reproduce the result in the paper

Hi, author
I wonder if you already release the all the code that can reproduce the result in the paper?
I could not find the Centroid_based Evaluation in this repo.

ID switch?

Hey @lannguyen0910 - How did the model performed with occlusions? Was there any ID switch , if so how frequent is it? I tried random erasing augmentation technique and use OSnet model , however when I tested it to track, the Id switches when there is change in pose of person. If the person is static, even with occlusion, the track ID remains constant.

Any help in this regard is appreciated.

Thank you for your great work.

Thanks
Ravi

Pretrained Models (on ImageNet) EfficientNet-v2 link not good

Hi thanks for sharing this repo
I try to load the EfficientNet-v2 ImageNet weights
but there is no net or model in the file

Pretrained Models link dead

can you share again? @lannguyen0910

An error occurs ' 'ValueError: too many values to unpack (expected 3)'

An error occurs ' 'ValueError: too many values to unpack (expected 3)'

When I run a code 'python train.py --config_file=resnet50_cuhk', the train starts successfully but at the end, after 50 epoch, an error occurs like this 'ValueError: too many values to unpack (expected 3)'
How can I solve this problem?

Download weights (ZIP): Permission denied

When running the colab jupyter notebook: https://colab.research.google.com/drive/1a-63lIx3mIU53C8aw1SU8l_GO_cackPX?usp=sharing#scrollTo=0bY292Jtao3K, got the permission issue

Download weights (ZIP)
[6]
2s
!gdown --id 12l1Z8qgVoLpjewwWwC1AN7_bK1_SIV_w
!gdown https://drive.google.com/uc?id=12l1Z8qgVoLpjewwWwC1AN7_bK1_SIV_w #efficientnetv2_market1501 weights
!gdown https://drive.google.com/uc?id=1gscms5dTajTmJ-0DrqOVTyjPZCi20Ku3 #efficientnetv2_cuhk03 weights
!gdown https://drive.google.com/uc?id=1hV9di9QRlKH1rMeqjlgc3HrpdsHuJKat #resnet50_market1501 weights
!gdown https://drive.google.com/uc?id=1Obd2jyg753Hiil86J0np51IQVWGp7AZf #resnet50_cuhk weights
Permission denied: https://drive.google.com/uc?id=12l1Z8qgVoLpjewwWwC1AN7_bK1_SIV_w
Maybe you need to change permission over 'Anyone with the link'?
Permission denied: https://drive.google.com/uc?id=12l1Z8qgVoLpjewwWwC1AN7_bK1_SIV_w
Maybe you need to change permission over 'Anyone with the link'?
Permission denied: https://drive.google.com/uc?id=1gscms5dTajTmJ-0DrqOVTyjPZCi20Ku3
Maybe you need to change permission over 'Anyone with the link'?
Permission denied: https://drive.google.com/uc?id=1hV9di9QRlKH1rMeqjlgc3HrpdsHuJKat
Maybe you need to change permission over 'Anyone with the link'?
Permission denied: https://drive.google.com/uc?id=1Obd2jyg753Hiil86J0np51IQVWGp7AZf
Maybe you need to change permission over 'Anyone with the link'?

Can we optimize the model for improving performance?

Need some inputs if we can convert the model file in ONNX or tensorrt.

Thanks in advcance.

Download links for Pretrained weights not working

First of all, thank you for the amazing work!

It seems that the links to download the ImageNet pre-trained models, datasets, and EfficientNet-v2 and Resnet50-IBN-A are not working, giving a 404 page not found error.

Train on custom dataset (with more no. of classes)

@lannguyen0910
How can I train the model on my custom dataset (number of classes: 2504) using the pre-train weights (number of classes: 751 for market).
I came across a method of removing the last or fully connected layer before loading the weights and then adding the fully connected layer (with desired num_classes) but wasn't able to implement it.
Could you please help in implementing the same?

Thanks and regards.

training parameter

Hi, I'm kazuki-can.
Thanks for your attractive work.
Since I'm going to write a paper about Computer Vision with this repository, I would like to know its training parameter.
I use the model file called "efficienetv2_imagenet.pt".

Thank you.

Inference on custom data

The way I understand this algorithm would work on custom data is as follows:

1. I run a tracker + object detector over my video and create "galleries" for each id identified by the tracker.
   • If I have an id switch because of occlusions or something similar I will have two galleries of the same person with two different ids
2. I save each crop identified by the object detector as a separate image in order to use as "queries"
3. Load a pretrained model (e.g efficientnet_v2_optimizer_45.pth) and run inference.py
   • This will calculate the features of each "gallery" I previously created as well as for each "query"
4. See to which of the galleries each "query" was "matched" and if it's more than one gallery assume there was an id switch there and that those galleries should belong to the same id.

Let me know if any of this makes sense or I should approach the problem differently. Thank you!

How to setup the folder structure?

1. Could you please share in which folder the datasets and pre-trained weights should be downloaded and placed?

2. Thanks for sharing the colab notebook (makes it very easy to test the code). Could you add the code to automatically download (wget) the datasets and pre-trained weights in the right folder in the notebook itself? Else it is quite confusing how to set up the directories to run the code.

Thank you for sharing the code with the community. @lannguyen0910

try this model on my data with pre-trained model by market1501

Hi, I am new to this model.
Thank you for this great work. This is quite interesting.
I have a question about evaluation on my data.
How can I use this pre-trained model?
I want to try with weights by market1501.
Can I write like below?
` gallery_feats = torch.load(PATH_TO_MY_GALLERY_DATA)
img_path = np.load(os.path.join(config.log_dir, config.img_path))
query_img = query_image
input = torch.unsqueeze(transform(query_img), 0)
input = input.to(config.device)
with torch.no_grad():
query_feat = model(input)

    dist_mat = cosine_similarity(query_feat, gallery_feats)
    indices = np.argsort(dist_mat, axis=1)
    visualizer(test_img, camid='mixed', top_k=10,img_size=config.image_size)`

Thank you.

Official repo released

Hi @lannguyen0910 ,

Thanks for your implementation of the reid model and for citing us as one of the papers.
Just wanted to share the news that we released our code along with trained model weights.
https://github.com/mikwieczorek/centroids-reid

Regards