The official repository for Benchmarks for Corruption Invariant Person Re-identification (NeurIPS 2021 Track on Datasets and Benchmarks), with exhaustive study on corruption invariant learning in single- and cross-modality ReID datasets, including Market-1501-C, CUHK03-C, MSMT17-C, SYSU-MM01-C, RegDB-C.

When deploying person re-identification (ReID) model in safety-critical applications, it is pivotal to understanding the robustness of the model against a diverse array of image corruptions. However, current evaluations of person ReID only consider the performance on clean datasets and ignore images in various corrupted scenarios. In this work, we comprehensively establish 5 ReID benchmarks for learning corruption invariant representation.

The benchmark will be maintained by the authors. We will get constant lectures about the new proposed ReID models and evaluate them under the CIL benchmark settings in time. Besides, we gladly take feedback to the CIL benchmark and welcome any contributions in terms of the new ReID models and corresponding evaluations. Please feel free to contact us, [email protected] .

TODO:

• other datasets configurations
• get started tutorial
• more detailed statistical evaluations
• checkpoints of the baseline models
• cross-modality preson Re-ID dataset, CUHK-PEDES
• vehicle ReID datasets, like VehicleID, VeRi-776, etc.

(Note: codebase from TransReID)

1. Install dependencies

• python=3.7.0
• pytorch=1.6.0
• torchvision=0.7.0
• timm=0.4.9
• albumentations=0.5.2
• imagecorruptions=1.1.2
• h5py=2.10.0
• cython=0.29.24
• yacs=0.1.6

2. Prepare dataset

Download the datasets, Market-1501, CUHK03, MSMT17. Set the root path of the dataset in congigs/Market/resnet_base.yml, DATASETS: ROOT_DIR: ('root'), or set it in scripts/train_market.sh, DATASETS.ROOT_DIR "('root')".

3. Train

Train a CIL model on Market-1501,

sh ./scripts/train_market.sh

4. Test

Test the CIL model on Market-1501,

sh ./scripts/eval_market.sh

The main code of corruption transform. (See contextual code in ./datasets/make_dataloader.py, line 59)

from imagecorruptions.corruptions import *

corruption_function = [gaussian_noise, shot_noise, impulse_noise, defocus_blur,
    glass_blur, motion_blur, zoom_blur, snow, frost, fog, brightness, contrast,
    elastic_transform, pixelate, jpeg_compression, speckle_noise,
    gaussian_blur, spatter, saturate, rain]
    
class corruption_transform(object):
    def __init__(self, level=0, type='all'):
        self.level = level
        self.type = type

    def __call__(self, img):
        if self.level > 0 and self.level < 6:
            level_idx = self.level
        else:
            level_idx = random.choice(range(1, 6))
        if self.type == 'all':
            corrupt_func = random.choice(corruption_function)
        else:
            func_name_list = [f.__name__ for f in corruption_function]
            corrupt_idx = func_name_list.index(self.type)
            corrupt_func = corruption_function[corrupt_idx]
        c_img = corrupt_func(img.copy(), severity=level_idx)
        img = Image.fromarray(np.uint8(c_img))
        return img

Evaluating corruption robustness can be realized on-the-fly by modifing the transform function uesed in test dataloader. (See details in ./datasets/make_dataloader.py, Line 266)

val_with_corruption_transforms = T.Compose([
    corruption_transform(0),
    T.Resize(cfg.INPUT.SIZE_TEST),
    T.ToTensor(),])

We introduce a rain corruption type, which is a common type of weather condition, but it is missed by the original corruption benchmark. (See details in ./datasets/make_dataloader.py, Line 27)

def rain(image, severity=1):
    if severity == 1:
        type = 'drizzle'
    elif severity == 2 or severity == 3:
        type = 'heavy'
    elif severity == 4 or severity == 5:
        type = 'torrential'
    blur_value = 2 + severity
    bright_value = -(0.05 + 0.05 * severity)
    rain = abm.Compose([
        abm.augmentations.transforms.RandomRain(rain_type=type, 
        blur_value=blur_value, brightness_coefficient=1, always_apply=True),
        abm.augmentations.transforms.RandomBrightness(limit=[bright_value, 
        bright_value], always_apply=True)])
    width, height = image.size
    if height <= 60:
        scale_factor = 65.0 / height
        new_size = (int(width * scale_factor), 65)
        image = image.resize(new_size)
    return rain(image=np.array(image))['image']

• Single-modality datasets

• Cross-modality datasets

Note: For RegDB dataset, Mode A and Mode B represent visible-to-thermal and thermal-to-visible experimental settings, respectively. And for SYSU-MM01 dataset, Mode A and Mode B represent all search and indoor search respectively. Note that we only corrupt RGB (visible) images in the corruption evaluation.

(Note: ranked by mAP on corrupted test set)

(Note: ranked by mAP on corrupted test set)

(Note: ranked by mAP on corrupted test set)

Kindly include a reference to this paper in your publications if it helps your research:

@misc{chen2021benchmarks,
    title={Benchmarks for Corruption Invariant Person Re-identification},
    author={Minghui Chen and Zhiqiang Wang and Feng Zheng},
    year={2021},
    eprint={2111.00880},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}