SBSGAN: Suppression of Inter-Domain Background Shift for Person Re-Identification

Abstract

Cross-domain person re-identification (re-ID) is challenging due to the bias between training and testing domains. We observe that if backgrounds in the training and testing datasets are very different, it dramatically introduces difficulties to extract robust pedestrian features, and thus compromises the cross-domain person re-ID performance. In this paper, we formulate such problems as a background shift problem. A Suppression of Background Shift Generative Adversarial Network (SBSGAN) is proposed to generate images with suppressed backgrounds. Unlike simply removing backgrounds using binary masks, SBSGAN allows the generator to decide whether pixels should be preserved or suppressed to reduce segmentation errors caused by noisy foreground masks. Additionally, we take ID-related cues, such as vehicles and companions into consideration. With high-quality generated images, a Densely Associated 2-Stream (DA-2S) network is introduced with Inter Stream Densely Connection (ISDC) modules to strengthen the complementarity of the generated data and ID-related cues. The experiments show that the proposed method achieves competitive performance on three re-ID datasets, i.e., Market-1501, DukeMTMC-reID, and CUHK03, under the crossdomain person re-ID scenario.

Domain Adaptive Person Re-Identification via Coupling Optimization

“…As DBC also predicts labels via unsupervised clustering, these results indicate that proposed GLO method can effectively improve the discriminative ability of learned features. It can also be observed that GLO also outperforms several domain adaptive methods such as PAUL [45] and DA_2S [14]. Specially, compared with ECN [60] that uses a source dataset, GLO still outperforms it by 2.7% in mAP on Market-1501.…”

“…Our method achieves Rank-1 accuracy of 77.4% on Market-1501, significantly outperforming recent BUC [25] and DBC [6] by 11.2% and 8.2%, respectively. It is worth noting that, our unsupervised training also outperforms several domain adaptive methods that use extra source domain for training, such as PAUL [45] and DA_2S [14].…”

“…Comparison on domain adaptive person ReID on Market-1501 and DukeMTMC-reID. Proposed DIM+GLO model is compared with recent state-of-the-art methods including TAUDL † [21], UTAL † [22], MAR [46], PAUL [45], CASCL [42], UDA [33], GPP [61], HHL [59], ECN [60], ATNet [27], CR_GAN [4], SSG ‡ [7], DA_2S [14], CAL * † [32], PDA-Net [23], and PAST ‡ [49]. The comparison results are summarized in Table 3.…”

“…Despite the significant success, domain adaptive person ReID is still a challenging task and there remain several open issues unexplored. Firstly, previous works commonly conduct the knowledge transfer in two stages, i.e., first transfer labeled images to the target domain with Generative Adversarial Networks (GANs), then train the ReID model [4,14,23,27] using transferred images. However, GANs could be hard to tune.The image generation is also challenging and sensitive to various factors like backgrounds, lighting, etc.. Secondly, person ReID models for the unlabeled target domain can be optimized by predicted labels.…”

“…Recently, many works have been proposed for domain adaptive person ReID task from different aspects. Some works targeting to narrow domain gaps in image space by GANs [4,14,23,27,40,59]. For example, Liu et al [27] propose an adaptive transfer network to effectively transfer images from label domain to unlabeled domain.…”