FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification
Authors: Yixiao Ge, Zhuowan Li, Haiyu Zhao, Guojun Yin, Shuai Yi, Xiaogang Wang, hongsheng Li
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our proposed FD-GAN achieves state-of-the-art performance on three person re ID datasets, which demonstrates that the effectiveness and robust feature distilling capability of the proposed FD-GAN. and Our method outperforms previous works in three widely-used re ID datasets, i.e. Market-1501 [5], CUHK03 [6] and Duke MTMC-re ID [7] datasets. and Table 1: Component analysis of the proposed FD-GAN on Market-1501 [5] and Duke MTMCre ID [7] datasets in terms of top-1 accuracy (%) and m AP (%) |
| Researcher Affiliation | Collaboration | 1CUHK-Sense Time Joint Laboratory, The Chinese University of Hong Kong 2Sense Time Research 3Johns Hopkins University 4University of Science and Technology of China |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is now available. https://github.com/yxgeee/FD-GAN |
| Open Datasets | Yes | In this paper, three datasets are used for performance evaluation, including Market-1501 [5], CUHK03 [6], and Duke MTMC-re ID [7]. |
| Dataset Splits | Yes | For each training mini-batch, it contains 128 person image pairs, with 32 of them belonging to same persons (positive pairs) and 96 of them belonging to different persons (negative pairs). All images are resized to 256 128. and The Market-1501 dataset [5] consists of 12,936 images of 751 identities for training and 19,281 images of 750 identities in the gallery set for testing. and We tune hyperparameters on the validation set of Market-1501 [5], and directly use the same hyperparameters for Duke MTMC-re ID [7] and CUHK03 [6] datasets. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like Res Net-50, Adam optimizer, and SGD, but does not provide specific version numbers for any software dependencies (e.g., deep learning frameworks or libraries). |
| Experiment Setup | Yes | For each training mini-batch, it contains 128 person image pairs, with 32 of them belonging to same persons (positive pairs) and 96 of them belonging to different persons (negative pairs). All images are resized to 256 128. The Gaussian bandwidth for obtaining pose landmark heat-map is uniformly sampled in [4, 6]. and The network is optimized by Stochastic Gradient Descent (SGD) with momentum 0.9. The initial learning rates are set to 0.01 for E and 0.1 for V , and they are decreased to 0.1 of their previous values every 40 epochs. The stage-I training process iterates for 80 epochs. and The initial learning rates for G, Did, Dpd are set as 10 3, 10 4, 10 2, respectively. Learning rates maintain the same for the first 50 epochs, and then gradually decrease to 0 in the following 50 epochs. The loss weights are set as λid = 0.1, λpd = 0.1, λr = 10, λsp = 1. |