This project tries to reproduce paper Beyond Part Models: Person Retrieval with Refined Part Pooling, and now is in processing.
In Market-1501:
| mAP (%) | Rank-1(%) | |
|---|---|---|
| Market-1501(paper) | 77.30 | 92.40 |
| Market-1501 | 74.03 | 89.43 |
usage: main.py [-h] [--params-filename PARAMS_FILENAME] [--use-gpu USE_GPU]
[--world-size WORLD_SIZE] [--dist-url DIST_URL]
[--dist-rank DIST_RANK] [--last-conv LAST_CONV]
[--batch-size BATCH_SIZE] [--num-workers NUM_WORKERS]
[--load-once LOAD_ONCE] [--epoch EPOCH] [--stage STAGE]
Person Re-Identification Reproduce
optional arguments:
-h, --help show help message and exit
--params-filename PARAMS_FILENAME
filename of model parameters.
--use-gpu USE_GPU set 1 if want to use GPU, otherwise 0. (default 1)
--world-size WORLD_SIZE
number of distributed processes. (default 1)
--dist-url DIST_URL the master-node's address and port
--dist-rank DIST_RANK
rank of distributed process. (default 0)
--last-conv LAST_CONV
whether contains last convolution layter. (default 1)
--batch-size BATCH_SIZE
training data batch size. (default 64)
--num-workers NUM_WORKERS
number of workers when loading data. (default 20)
--load-once LOAD_ONCE
load all of data at once. (default 0)
--epoch EPOCH number of epochs. (default 60)
--stage STAGE running stage. train, test or all. (default train)