Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification

Abstract

Person re-identification (re-ID) models trained on one domain often fail to generalize well to another. In our attempt, we present a "learning via translation" framework. In the baseline, we translate the labeled images from source to target domain in an unsupervised manner. We then train re-ID models with the translated images by supervised methods. Yet, being an essential part of this framework, unsupervised image-image translation suffers from the information loss of source-domain labels during translation.Our motivation is two-fold. First, for each image, the discriminative cues contained in its ID label should be maintained after translation. Second, given the fact that two domains have entirely different persons, a translated image should be dissimilar to any of the target IDs. To this end, we propose to preserve two types of unsupervised similarities, 1) self-similarity of an image before and after translation, and 2) domain-dissimilarity of a translated source image and a target image. Both constraints are implemented in the similarity preserving generative adversarial network (SPGAN) which consists of an Siamese network and a Cy-cleGAN. Through domain adaptation experiment, we show that images generated by SPGAN are more suitable for domain adaptation and yield consistent and competitive re-ID accuracy on two large-scale datasets.

Cross-Dataset Person Re-Identification via Unsupervised Pose Disentanglement and Adaptation

“…To better exploit and adapt visual information across data domains, methods based on domain adaptation [8,24] have been utilized [12,14,29,35,49,61]. However, since the identities, viewpoints, body poses and background clutter can be very different across datasets, plus no label supervision is available at the target domain, the performance gains might be limited.…”

“…In Table 1, we compare our proposed model with the use of Bag-of-Words (BoW) [58] for matching (i.e., no transfer), four unsupervised re-ID approaches, including UMDL [42], PUL [15], CAMEL [54] and TAUDL [29], and seven cross-dataset re-ID methods, including PTGAN [51], SPGAN [12], TJ-AIDL [49], MMFA [35], HHL [61], CFSM [3] and ARN [32]. From this table, we see that our model achieved very promising [12] and HHL [61], we note that our model is able to generate cross-domain images conditioned on various poses rather than few camera styles. Compared to MMFA [35], our model further disentangles the pose information and learns a pose invariant cross-domain latent space.…”

“…In Figure 4, we additionally consider the crossdataset re-ID appoach of SPGAN [13] and the posedisentanglement re-ID method of FD-GAN [18]. We see that, since SPGAN performed style transfer for synthesizing cross-domain images, pose variants cannot be exploited in the target domain.…”

“…Since it might be computationally expensive to collect identity labels for the dataset of interest, one popular solution is to utilize an additional yet distinct source-domain dataset. This dataset contains fully labeled images (but Figure 1: Existing cross-dataset re-ID methods like [12] perform style transfer followed by feature extraction for re-ID, which might limit image variants to be observed. We choose to perform pose disentanglement and adaption with domain-invariant features jointly learned, alleviating the above issue with improved image representation.…”

“…Several methods for cross-dataset re-ID have been proposed [13,15,42,49,54,58,61]. For example, Deng et al [13] employ CycleGAN to covert labeled images from source to target domains, followed by performing re-ID at the target domain. Similarly, Zhong et al [61] utilize Star-GAN [11] to learn camera invariance and domain connectedness simultaneously.…”