Deep Person Reid Versions Save

Major changes

• https://github.com/KaiyangZhou/deep-person-reid/commit/1e9d466f42256cc451f6f73761c298cabbcd0b39
• https://github.com/KaiyangZhou/deep-person-reid/commit/4a033659b0330bcbd25dc1cc344cf26ddd69ac73

Removed prid450s.

This version should be stable.

Bug fix:

• https://github.com/KaiyangZhou/deep-person-reid/commit/235348f67248a9d27d5c9bcafcabcfe5cbf61cb3: return ImageDataset or VideoDataset rather than Dataset.

See the documentation https://kaiyangzhou.github.io/deep-person-reid/.

Major updates:

• Model codes such as resnet.py and densenet.py keep the original style for easier modification.
• Generalize CrossEntropyLableSmooth to CrossEntropyLoss. --label-smooth should be called explicitly in order to add the label smoothing regularizer to the cross entropy loss.
• Add support to multi-dataset training. Datasets are specified by the arguments -s and -t, which refer to source datasets and target datasets, respectively. Both can take multiple strings delimited by space. For example, say you wanna train a model using Market1501+DukeMTMC-reID, just set -s market1501 dukemtmcreid. If you wanna test on multiple datasets, you can do -t market1501 dukemtmcreid cuhk03 msmt17.
• Arguments are unified in args.py.
• Dataloaders are wrapped into two classes, which are ImageDataManager and VideoDataManager (see data_manager.py). A datamanger is initialized by dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args)) where image_dataset_kwargs() is implemented in args.py. Therefore, when new arguments are added to the data manager, you don't need to exhausively change everywhere in the code. What you need to update are (1) add new arguments in args.py and (2) update the input arguments in data_manager.py.
• BENCHMARK is replaced with MODEL_ZOO where model weights and training scripts can be downloaded.

• Added --lambda-xent and --lambda-htri in xxx_xent_htri.py, which can balance between cross entropy loss and hard mining triplet loss.
• Divided losses into separate files for easier extension.
• Moved major codes to the folder torchreid/ (such structuring will be maintained).
• Automated the download of dukemtmcreid and dukemtmcvidreid.

• Added --load-weights (weights that don't match in size will be discarded, e.g. old classification layer).
• Updated dukemtmcvidreid naming, old/new formats are supported.
• Added --vis-ranked-res and reidtools.py, allowing ranked images to be visualized.

Note: --use-lmdb is postponed.

• Code structure is re-organized, which is easier to read and modify
• data_manager is split into different scripts.
• Added --fixbase option, allowing randomly initialized classifier layer to be trained while keeping base network frozen.
• Rewrote sampler for triplet loss, which allows most of the training images to be covered during a training epoch.
• Added requirements.txt, which eases environment configuration.
• Instructions on data preparation and benchmarks are separated into different files.

To be done:

• --lmdb is under development.

added cython-based evaluation.

• multi-GPU training.
• both image-based and video-based reid.
• unified interface for different reid models.
• easy dataset preparation.
• end-to-end training and evaluation.
• standard dataset splits used by most papers.
• download of trained models.