Deep Person Re-identification Introduction

Introduction to person re-identification using deep learning models. It includes experiments about cross-domain issue.
Experiments are based on open-source github code: github.com/KaiyangZhou/deep-p....
Backbones used are based on paper:
Learning Generalisable Omni-Scale Representations for Person Re-Identification. Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, Tao Xiang.
Tricks used are based on paper:
Bag of tricks and A strong baseline for deep person re-identification. Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, Wei Jiang.

[@sharath_9246]

Thanks For the Video🙏🏻 very much appreciated!!
Metrics Explanation is spot on

[@codingcat3274]

Glad you enjoyed it!

[@suenosn562]

great presentation, i would like to see more videos on prson re-id if possible, thank you in advance

[@songpandy9590]

Great Video. Thank you.

[@hamidhabibsyed9286]

Can you make some further videos on person reidentification?

[@slann2k]

This video is super helpful! could you share some of the code you used to run the experiments?

[@codingcat3274]

I mainly use this open source code: github.com/KaiyangZhou/deep-person-reid. It is pretty handy.

[@prhmma]

thank you for the resource shared

[@codingcat3274]

Glad it was helpful!

[@abhishekomi1573]

Helpful video for me.. Thanks 😀

[@codingcat3274]

Glad to hear it :)

[@editssinfinity5355]

Hi, thanks for your explanation. This is really awesome. I have go through the GitHub repo. I didn't find any way to infer reid on real time i.e. on a video. There is only I find is training and evaluation on some already designed datasets. Can you please help me out with this?

[@codingcat3274]

Hi, if you use reid every single frame, it can hardly be real time. To achieve real-time performance, usually it is a combination of reid and tracking, as tracking is much faster than reid, while reid is more accurate.

[@barakullah949]

i want to ask that this torchreid model does perform tracking as well with the reidentification or it just performs the reidentification?

[@codingcat3274]

It only performs person reid. You could use some naive tracking frame to frame. But when there is occlusion, it is difficult for both tracking and person reid.

[@hamidhabibsyed9286]

👍Thank you

[@codingcat3274]

Glad it helps :D

[@vipingautam9501]

Great presentation, may I know how much time does it take for inference on image pairs?
I am considering this with Yolo+Tracker(Possibly opencv one).

[@codingcat3274]

Depends on the gallery size if you want to find all image pairs, but for getting the embedding itself, it takes about 200 ms for each image.

[@vipingautam9501]

@@codingcat3274 That's very slow, because I was thinking of using it for realtime. Example: when a person leaves one FOV and moves to another FOV and performing re-id in such cases..do you think it would be application in such cases?

[@codingcat3274]

Yes, I think that is what person reid is designed for, but to achieve real time, there might be sacrifice on the model accuracy and you might use some approximate search algorithm.

[@vipingautam9501]

@@codingcat3274 Thanks for your insights on this, would you mind recommending some good algorithms for aproximate search? I am working on this as a part of my Thesis I have to build a system

[@codingcat3274]

You can check out ANN for searching.

[@abdelhasibnaamaoui7238]

thanks

[@codingcat3274]

No problem!

[@siddharthrana512]

how you make multi camera to communicate so that each camera predict the same id.can you please explain it

[@codingcat3274]

All images from multi cameras are stored in the gallery, and when we query a person, we try to find the same person in the gallery with different camera id.

[@biancoluke]

what is the possibility to provide your snipped demo? Thanks in advance.

[@codingcat3274]

I can send you the ranking visualization results via email. But I highly recommend to run the torchreid repo since there are so many combinations of various datasets / networks / hyperparameters.

[@SAINIVEDH]

Why no one uses OSNET. Isn't that architecture specially designed for Person re-id tasks ??.

[@codingcat3274]

I find that when it comes to cross domain, osnet_ibn and osnet_ain work quite well compared with other models. But in general, person reid is just a more difficult problem to solve.

[@ThanhNguyen-dq2eh]

Thanks for the presentation. Is there any way that I can have your slides?

[@codingcat3274]

Sure, I can send you via email.

[@ThanhNguyen-dq2eh]

@@codingcat3274 here is my email thanh1705 at hotmail dot com

[@gdsu234]

@@codingcat3274 Hey! Thanks a ton for the video. Could you share the slides with me too? My email is [e-mail]

[@hamidhabibsyed9286]

I am interested in '' Person Reidentification '' research, so if you can guide?

[@codingcat3274]

The topics were randomly chosen. But I could prepare more videos on Person Reid.

[@hamidhabibsyed9286]

@@codingcat3274 I will be waiting for it! 😊

[@suenosn562]

@@codingcat3274 hello, i'm also interested in person re-id, i'm waiting for your new

[@GTARobotics]

Thank you for this amazing presentation!
I would like to integrate person re-identification algorithm in OSSDC VisionAI platform, see my latest videos with demos.
Do you have an example on how to run this on webcam video source?

[@codingcat3274]

I checked your demos. They are really cool! You can have a gallery dataset, and preprocess the images to get embeddings for each person id with the person reid model. Then in your query video, each person is detected by person detector first, compute the embedding according to the person reid model, and compare this query embedding to gallery embeddings. And you can re-identify the person by finding the shortest distance in gallery images and a threshold. It is very similar to face recognition.

[@GTARobotics]

@@codingcat3274 Thanks!
I'll try what you suggested.

[@SAINIVEDH]

@@codingcat3274 So for huge shopping malls/public places wouldn't this approach be too slow as there will be thousands of gallery images and we need to process all those for each frame

[@codingcat3274]

@@SAINIVEDH Right, usually, person reid model won't be used for every single frame, usually it is a combination of person reid + tracking. For example, every 10 frames, use person reid model once. Or when there is great occlusion or the person quality score is low, use reid model.

[@thinkman2137]

Can you run Bag of Tricks and A Strong ReID Baseline source code?

[@codingcat3274]

Yes, I could, the results are based on running their code.

[@thinkman2137]

@@codingcat3274 can I get your weight, 120 epoch too much

[@codingcat3274]

That depends on your training set, the number of epoches varies according to your training set size. For me, I usually start with something around 40 and then check the curves, if it already converges, try to make it 1/2. But I will keep it bigger than 10.

[@thinkman2137]

@@codingcat3274 how your mAP, is it bigger than 80%?

[@thinkman2137]

@@codingcat3274 my train 25 epoch 71.4 mAP, second train 58 eooch 70 mAP. Look not good