xiaoccer/reid-pcb_rpp

Beyond Part Models: Person Retrieval with Refined Part Pooling

ReID-PCB_RPP

This repository tries to implement the paper:Beyond Part Models: Person Retrieval with Refined Part Pooling and I almost follow the training details at the origin paper.

Market-1501 Results

• The result of PCB is little higher than paper.
• The result of PCB+RPP is little lower than paper.
• I think if we do more augmentation with the training data, the performance will be better.

	Feature	mAP(%)	Rank-1(%)
PCB(paper)	G	77.4	92.3
PCB(paper)	H	77.3	92.4
PCB(ours)	G	78.6	92.7
PCB(ours)	H	78.5	92.1
PCB+RPP(paper)	G	81.6	93.8
PCB+RPP(paper)	H	81.0	93.1
PCB+RPP(ours)	G	80.7	92.8
PCB+RPP(ours)	H	79.8	92.4

Prerequisites

• Python 2.7 or 3.5
• Pytorch 0.4
• GPU Memory

Getting started

Dataset

• Dwonload Market1501 Dataset and extract the files to the current folder
• Change the download_path in the prepare.py and run python prepare.py to prepare the dataset

Train

Train the model:

python train.py --gpu_ids 0,1 --batchsize 64 --data_dir your_data_path --save_dir your_model_save_path --RPP True

• --gpu_ids: single GPU or mutil-GPUs
• --batchsize: batch size
• --data_dir: the path of the preparing data
• --save_dir: the path to save the training model
• --RPP: whether to train with RPP I train the model on the two GPUs with 64 batchsize

Test

Test the model(extract the features):

python test.py --gpu_ids 0 --which_epoch select the model --stage PCB or full --RPP True or False --feature_H True or False

• --gpu_ids: just single GPU
• --which_epoch: select the i-th model
• --stage: select the training model : PCB or full(PCB+RPP)
• --RPP: if you choose the PCB model, RPP to False; if you choose the full model, RPP to True
• --feature_H: whether to extract the low-dims features

Evaluate

Evaluate the model

• Evaluate on GPU:

python evaluate_gpu.py --gpu_ids 0 --reslut_mat the path of  the features mat

• Evaluate on CPU:

python evaluate_gpu.py --reslut_mat the path of  the features mat

Reference resources

Thanks to the layumi/Person_reID_baseline_pytorch. This repository only implements the part of PCB, I make some modifications on it and then add the RPP part.

reid-pcb_rpp's Issues

Question about input data normalization [mean,std] value where come from

ReID-PCB_RPP/train.py

Line 74 in 00462d0

transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])

and the val data in line 79
is same ?

how different from https://github.com/layumi/Person_reID_baseline_pytorch

Hi,
Thank you for sharing your work.
Not an issue but I just wonder what is the most important difference of your project from layumi's because he gets only ~89 and you improve it to ~92 (PCB part only).

Errors accur when running test.py

File "test.py", line 157, in
model = load_network(model_structure)
File "test.py", line 80, in load_network
network.load_state_dict(torch.load(save_path))
File "/home/lingck/.local/lib/python3.7/site-packages/torch/nn/modules/module.py", line 777, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for PCB:
Missing key(s) in state_dict: "avgpool.add_block.0.weight", "avgpool.norm_block.0.weight", "avgpool.norm_block.0.bias", "avgpool.norm_block.0.running_mean", "avgpool.norm_block.0.running_var".

do you have some guide for ｃustomized images for usage?

thanks for your great work. but do you have some guide for the customized images both have "gallery" and "query" of usage? it may helpful. thanks a lot.

the running_loss, what mean, i could not understand, is there any formula, thank you in advance

ReID-PCB_RPP/train.py

Line 174 in 00462d0

running_loss += loss.item() * inputs.size(0)

nice project！ clear structure and easy to read！

代码看起来非常舒适

train loss equals 0.0000 ,what is the problem?

when i train the model, only change the dataset path of Market-1501, the train loss equals 0.0000 ,what is the problem? thanks for your answer.

{'train': 12936, 'val': 19732} ['bounding_box_train']
Epoch 0/59

train Loss: 0.0000 Acc: 1.0000

Epoch 1/59

train Loss: 0.0000 Acc: 1.0000

Epoch 2/59

--RPP True 出错

您好，我运行您的代码出现了下面这个问题
为什么我使用这个--RPP True 命令的时候，提示True是无效？
error: unrecognized arguments: True
不使用--RPP True程序运行正常
请问这是什么问题？

very nice code, +1, learning from yours