wangguanan / pytorch-person-reid-baseline-pcb-beyond-part-models

A strong implementation of PCB (Beyond Part Models), outperforming all existing implementations.

Pytorch-Person-ReID-PCB-Beyond-Part-Models

• A Strong Implementation of PCB (Beyond Part Models, ECCV2018) on Market-1501 and DukeMTMC-reID datasets.

• We support:

  • A strong PCB implementation which outperforms most existing implementations.
  • A simple and clear implementation, and end-to-end training and evaluation.

News

• We re-write a strong Re-ID baseline Bag of Tricks (BoT) with a more simple and clear implementation, which is more friendly with researchers and newers. Our code can be found here. BoT outperforms PCB by using only the global feature. Our implementation of BoT achieves the same performance with the offical one.

Dependencies

• Anaconda (Python 3.7)
• PyTorch 0.4.0
• GPU Memory >= 20G (we use 2 GTX1080ti)
• Memory >= 20G

Dataset Preparation

• Market-1501 Dataset and DukeMTMC-reID Dataset
• Download and extract both anywhere

Train and Test

python main.py --market_path market_path --duke_path duke_path

Experiments

1. Settings

• We conduct our experiments on 2 GTX1080ti GPUs

2. Results

Implementations	market2market	duke2duke	market2duke	duke2market
PCB w/ REA (Ours)	0.939 (0.832) <model.pth>	0.856 (0.753) <model.pth>	0.384 (0.237)	0.555 (0.285)
PCB (Ours)	0.934 (0.809)	0.867 (0.746)	0.440(0.265)	0.592 (0.308)
PCB (layumi)	0.926 (0.774)	0.642 (0.439)	-	-
PCB (huanghoujing)	0.928 (0.785)	0.845 (0.700)	-	-
PCB (Xiaoccer)	0.927 (0.796)	-	-	-
PCB (Paper)	0.924 (0.773)	0.819 (0.653)	-	-

Contacts

If you have any question about the project, please feel free to contact with me.

E-mail: [email protected]

pytorch-person-reid-baseline-pcb-beyond-part-models's Issues

pcb

Have you ever tried to add an RPP section to your newest PCB module?Is it better than the original paper?

Result lower far than reported?

120epoch for p-k samples may be not enough? what is the correct training params?

visualization map

In the published PCB paper, the author analyzed the highlight area maps of different branches (e.g. Fig.1 and Fig.7), do you know how to calculate these maps?

Validation dataset are need?

Hi, Guan'an Wang

My English is very bad. I only use Google Translate to communicate. If I can communicate in Chinese, it’s best.

I have a ReID task problem for several days,
I think your code only uses train and test data sets,
No validation dataset is used.

I would like to ask your thoughts.
For ReID tasks, do you think validation is necessary?

According to the ReID tasks I have watched in the past few months,
some people are useful and some are not useful.
Hard to have a definite idea

I think for machine learning problems, validation is always important, since you need to make sure that your model is gaining generalization power instead of overfitting the test set. But since some areas still have very limited open data (still including ReID I guess?), the datasets used by academia don't have enough data for validation. With just a few validation data, the validation error might be pretty random and would not be very meaningful. So for research propose, I think it's fine to use just the training set and run massive experiments on serveral datasets for method comparison. But for enginners who aim to build a reliable model in the wild, it's necessary to collect a large amount of data and train the model with validation.

Have you ever thought about why you didn’t use the validation dataset when you wrote the program?

what is the purpose of embeedings_list in models.py?

hey, i found the embedding feautres are not used in your code, what is the purpose of embeding_list in models.py then?
thank you !

Abouts the results

Time: 2020-08-05 20:51:44; Test Dataset: market,
mAP: 0.4618259404352151
Rank: [0.7131829 0.7891924 0.82214964 0.84649644 0.86252969 0.87618765
0.88539192 0.89103325 0.89786223 0.90498812 0.90855107 0.91270784
0.9165677 0.92072447 0.92309976 0.92666271 0.93022565 0.93289786
0.93467933 0.93735154 0.9388361 0.94091449 0.94239905 0.94447743
0.945962 0.94685273 0.94774347 0.94833729 0.94982185 0.9510095
0.95190024 0.95249406 0.95308789 0.95368171 0.95427553 0.95486936
0.9557601 0.95665083 0.95724466 0.9584323 0.95932304 0.95961995
0.9608076 0.96110451 0.96140143 0.96140143 0.96169834 0.96169834
0.96258907 0.96347981 0.96347981 0.96407363 0.96407363 0.96466746
0.96466746 0.96526128 0.96555819 0.96585511 0.96585511 0.96615202
0.96674584 0.96674584 0.96704276 0.96704276 0.96733967 0.96793349
0.9682304 0.96852732 0.96882423 0.96912114 0.96941805 0.96971496
0.97030879 0.9706057 0.97149644 0.97179335 0.97209026 0.97238717
0.97238717 0.97238717 0.972981 0.972981 0.972981 0.972981
0.97327791 0.97416865 0.97416865 0.97416865 0.97416865 0.97416865
0.97446556 0.97446556 0.97446556 0.97505938 0.97505938 0.97535629
0.97535629 0.97535629 0.97535629 0.97565321 0.97595012 0.97595012
0.97624703 0.97624703 0.97624703 0.97624703 0.97624703 0.97624703
0.97624703 0.97654394 0.97684086 0.97684086 0.97684086 0.97684086
0.97713777 0.97743468 0.97743468 0.97773159 0.9780285 0.9780285
0.97832542 0.97891924 0.97921615 0.97951306 0.97951306 0.97951306
0.97951306 0.98010689 0.98010689 0.98010689 0.9804038 0.98070071
0.98070071 0.98070071 0.98070071 0.98070071 0.98099762 0.98099762
0.98099762 0.98129454 0.98129454 0.98129454 0.98129454 0.98129454
0.98129454 0.98159145 0.98159145 0.98159145 0.98159145 0.98159145]

I got the above results. I don't understand why I get 150 ranks after 120 epochs. What do they represent?