Feature-Level Adversarial Attacks and Ranking Disruption for Visible-Infrared Person Re-identification

Authors: Xi Yang, Huanling Liu, De Cheng, Nannan Wang, Xinbo Gao

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two VIRe ID benchmarks (i.e., SYSU-MM01, Reg DB) and different systems validate the effectiveness of our method. and 4 Experiments
Researcher Affiliation Academia Xi Yang1, Huanling liu1, De Cheng1 , Nannan Wang1, Xinbo Gao2 1Xidian University, 2Chongqing University of Posts and Telecommunications yangx@xidian.edu.cn,huanlingliu@stu.xidian.edu.cn, dcheng@xidian.edu.cn nnwang@xidian.edu.cn, gaoxb@cqupt.edu.cn
Pseudocode No The paper describes the method using text and figures, but does not include a formal pseudocode or algorithm block.
Open Source Code No Justification: We will consider making the code details publicly available in the future.
Open Datasets Yes SYSU-MM01[1] is a cross-modality pedestrian re-identification dataset proposed in 2017. and The Reg DB [43]dataset comprises 412 individuals, with each person having 20 images, including 10 visible images and 10 infrared images.
Dataset Splits No During training, we randomly sampled 16 identities, each with 4 images, to form a minibatch of size 64. The paper describes datasets and training/testing phases but does not explicitly provide percentages or counts for train/validation/test splits.
Hardware Specification Yes The experiments are conducted on an NVIDIA Ge Force 3090 GPU with Pytorch.
Software Dependencies No The experiments are conducted on an NVIDIA Ge Force 3090 GPU with Pytorch. Pytorch is mentioned without a version number, and no other specific software dependencies with versions are provided.
Experiment Setup Yes During training, we randomly sampled 16 identities, each with 4 images, to form a minibatch of size 64. Pedestrian images are resized to 288 144. Data augmentation included random horizontal flipping and random erasing with a probability of 0.5. We optimize using the stochastic gradient descent (SGD) optimizer, with a weight decay set to 0.0005 and a momentum parameter set to 0.9. The initial learning rate for both datasets was set to 0.1, and it was decayed by a factor of 0.1 at the 20th and 50th epochs, respectively. Finally, the margin α in the auxiliary quadruplet adversarial loss function is set to 0.2. We adopt a warm-up learning rate scheme, with a total of 60 training epochs.