Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Exploring the Camera Bias of Person Re-identification
Authors: Myungseo Song, Jin-Woo Park, Jong-Seok Lee
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically investigate the camera bias of person re-identification (Re ID) models. Previously, camera-aware methods have been proposed to address this issue, but they are largely confined to training domains of the models. We measure the camera bias of Re ID models on unseen domains and reveal that camera bias becomes more pronounced under data distribution shifts. ... Furthermore, we validate its generalizability across various models and benchmarks, highlighting its potential as a simple yet effective test-time postprocessing method for Re ID. |
| Researcher Affiliation | Collaboration | Myungseo Song & Jin-Woo Park m Ay-I Inc. Seoul, Korea EMAIL Jong-Seok Lee Yonsei University Seoul, Korea EMAIL |
| Pseudocode | Yes | Algorithm 1 Unsupervised learning algorithm for Person Re ID with simple modificaitons Require: Initialized backbone encoder fθ and training samples with camera labels X for n in [1, num_epochs] do Extract features F from X by fθ. (1) Debiased pseudo labeling: Transform F to ˆ F by applying the camera-specific feature normalization. Generate pseudo labels by clustering ˆ F. (2) Discarding biased clusters: Collect the images belong to the clusters of single camera as B. Reconstruct training images by X = X B. Prepare for training iterations (e.g., initialization of feature memory). for i in [1, num_iterations] do Sample a mini-batch from the reconstructed data X . Compute loss (e.g., contrastive loss). Update the encoder fθ. Update auxiliary modules (e.g., update of feature memory). end for end for |
| Open Source Code | Yes | Reproducibility The code for reproducing the experiment results is provided in the supplementary material. |
| Open Datasets | Yes | We observe that existing Re ID models exhibit a large camera bias for unseen domain data. For example, Figure 1 describes the feature distance distributions between samples of a camera-aware model (Cho et al., 2022) trained on the Market-1501 (Zheng et al., 2015) dataset, using samples from the MSMT17 (Wei et al., 2018) dataset. ... The results on MSMT17, Market-1501, CUHK03-NP (Zhong et al., 2017a), and Person X (Sun & Zheng, 2019) are shown in Table 1... |
| Dataset Splits | Yes | All models except ISR (Dou et al., 2023) are trained on MSMT17, hence the other datasets are unseen domains for them. ... We constructed a dataset of 7500 samples by randomly selecting 500 identities from Market-1501, where each identity has 5 samples per camera with 3 cameras. ... We constructed five datasets of 11821 samples of 1041 identities from MSMT17, where the maximum numbers of cameras per identity are different. |
| Hardware Specification | Yes | The models are trained on a H100 GPU with batch size 256 and 100 training epochs, with DBSCAN (Ester et al., 1996) to obtain pseudo labels. |
| Software Dependencies | No | The paper mentions using Media Pipe (Lugaresi et al., 2019) for body keypoint extraction, but does not specify a version number for it or any other software component used in the experiments. |
| Experiment Setup | Yes | The models are trained on a H100 GPU with batch size 256 and 100 training epochs, with DBSCAN (Ester et al., 1996) to obtain pseudo labels. ... Table 12: Training results of CC with the varying ϵ parameter of the DBSCAN algorithm. |