i want to know how to train on my own dataset,so please help me,thanks about person-reid-triplet-loss-baseline

[huanghoujing]

Thank you for your attention!

Least Modification for Your Usage

You can partition your data into trainval set and test set. Then, in your trainval set, you can separate some ids as val set. In this way, you have train, val, test sets.

For partitioning the test set or val set into query and gallery sets, you can use method partition_train_val_set in ./tri_loss/utils/dataset_utils.py.

If you don't want to rename your images as I do, you have to modify the parse_im_name method in ./tri_loss/utils/dataset_utils.py. Otherwise, you should rename your images to be with format new_im_name_tmpl = '{:08d}_{:04d}_{:08d}.jpg' (in ./tri_loss/utils/dataset_utils.py).

You have to save a dictionary with the following keys to a pickle file named partitions.pkl:

• trainval_im_names, the list of trainval image names
• trainval_ids2labels, a dictionary mapping trainval ids to labels
• train_im_names, the list of train image names
• train_ids2labels, a dictionary mapping train ids to labels
• val_im_names, the list of val image names
• val_marks, a list with the same length as val_im_names, containing 0 (denoting query image), 1 (denoting gallery image) or 2 (denoting multi-query image)
• test_im_names, the list of test image names
• test_marks, a list with the same length as test_im_names, containing 0 (denoting query image), 1 (denoting gallery image) or 2 (denoting multi-query image)

Usually, you wouldn't have multi-query set.

Place all images inside the same directory. Then add a new entry to tri_loss/dataset/__init__.py, as follows

elif name == YOUR-DATASET-NAME:
    im_dir = ospeu(YOUR-DATASET-IMAGE-DIRECTORY)
    partition_file = ospeu(YOUR-DATASET-PARTITION-PICKLE-FILE)

In the training script ./script/experiment/train.py, add your dataset name to this line, as follows

parser.add_argument('--dataset', type=str, default='market1501',
                        choices=['market1501', 'cuhk03', 'duke', 'combined', YOUR-DATASET-NAME])

Then the code should be running well on your dataset.

Compatibility with Provided Datasets

Since all datasets in tri_loss/dataset/__init__.py use the same parse_im_name method specified in ./tri_loss/utils/dataset_utils.py, you have to rename your image names to this format so that the code can run on all these datasets.

For renaming your image names to required format and moving all images to a directory, you can use method move_ims in tri_loss/utils/dataset_utils.py. Note that parse_im_name in def move_ims(ori_im_paths, new_im_dir, parse_im_name, new_im_name_tmpl) should be corresponding to your original image names.

For Extremely Large Dataset

If your dataset has extremely large number of images, e.g. hundreds of thousands, it may not be wise to place all images inside a single dataset, because retrieving an item from a directory containing e.g. millions of items may be time consuming. In this case, you can keep your original directory layout and provide relative image paths, e.g. sub-dir/image-name.jpg, to the code where image names are used. Some minor modification may also be needed in some places.

[JJJackssson]

thank you very much for your timely help @huanghoujing

[surajjunni]

Actually I am a beginner in computer vision.I was going through your post.I could not understand how to use my own datasets for align re-id.Can you help me in this?