Real-world visual search systems involve deployments on multiple platforms with different computing and storage resources. Deploying a unified model that suits the minimal-constrain platforms leads to limited accuracy. It is expected to deploy models with different capacities adapting to the resource constraints, which requires features extracted by these models to be aligned in the metric space. The method to achieve feature alignments is called “compatible learning”. Existing research mainly focuses on the one-to-one compatible paradigm, which is limited in learning compatibility among multiple models. We propose a Switchable representation learning Framework with Self-Compatibility (SFSC). SFSC generates a series of compatible sub-models with different capacities through one training process. The optimization of sub-models faces gradients conflict, and we mitigate this problem from the perspective of the magnitude and direction. We adjust the priorities of sub-models dynamically through uncertainty estimation to co-optimize sub-models properly. Besides, the gradients with conflicting directions are projected to avoid mutual interference. SFSC achieves state-of-the-art performance on the evaluated datasets.

Removing the undesired moiré patterns from images capturing the contents displayed on screens is of increasing research interest, as the need for recording and sharing the instant information conveyed by the screens is growing. Previous demoiréing methods provide limited investigations into the formation process of moiré patterns to exploit moiré-specific priors for guiding the learning of demoiréing models. In this paper, we investigate the moiré pattern formation process from the perspective of signal aliasing, and correspondingly propose a coarse-to-fine disentangling demoiréing framework. In this framework, we first disentangle the moiré pattern layer and the clean image with alleviated ill-posedness based on the derivation of our moiré image formation model. Then we refine the demoiréing results exploiting both the frequency domain features and edge attention, considering moiré patterns' property on spectrum distribution and edge intensity revealed in our aliasing based analysis. Experiments on several datasets show that the proposed method performs favorably against state-of-the-art methods. Besides, the proposed method is validated to adapt well to different data sources and scales, especially on the high-resolution moiré images.

Visual retrieval system faces frequent model update and deployment. It is a heavy workload to re-extract features of the whole database every time. Feature compatibility enables the learned new visual features to be directly compared with the old features stored in the database. In this way, when updating the deployed model, we can bypass the inflexible and time-consuming feature re-extraction process. However, the old feature space that needs to be compatible is not ideal and faces outlier samples. Besides, the new and old models may be supervised by different losses, which will further causes distribution discrepancy problem between these two feature spaces. In this article, we propose a global optimization Dual-Tuning method to obtain feature compatibility against different networks and losses. A feature-level prototype loss is proposed to explicitly align two types of embedding features, by transferring global prototype information. Furthermore, we design a component-level mutual structural regularization to implicitly optimize the feature intrinsic structure. Experiments are conducted on six datasets, including person ReID datasets, face recognition datasets, and million-scale ImageNet and Place365. Experimental results demonstrate that our Dual-Tuning is able to obtain feature compatibility without sacrificing performance.

In object re-identification (ReID), the development of deep learning techniques often involves model updates and deployment. It is unbearable to re-embedding and re-index with the system suspended when deploying new models. Therefore, backward-compatible representation is proposed to enable ``new'' features to be compared with ``old'' features directly, which means that the database is active when there are both ``new'' and ``old'' features in it. Thus we can scroll-refresh the database or even do nothing on the database to update. The existing backward-compatible methods either require a strong overlap between old and new training data or simply conduct constraints at the instance level. Thus they are difficult in handling complicated cluster structures and are limited in eliminating the impact of outliers in old embeddings, resulting in a risk of damaging the discriminative capability of new features. In this work, we propose a Neighborhood Consensus Contrastive Learning (NCCL) method. With no assumptions about the new training data, we estimate the sub-cluster structures of old embeddings. A new embedding is constrained with multiple old embeddings in both embedding space and discrimination space at the sub-class level. The effect of outliers diminished, as the multiple samples serve as ``mean teachers''. Besides, we propose a scheme to filter the old embeddings with low credibility, further improving the compatibility robustness. Our method ensures the compatibility without impairing the accuracy of the new model. It can even improve the new model's accuracy in most scenarios.

Part-level features obtained by uniformly partitioning have attracted much attention in person re-identification. However, standard uniform part partitions may lead to within-part inconsistency across different samples, as shown in Figure 1. Attention mechanisms (e.g., refined part pooling) have been proposed to refine part division with enhanced consistency. Unfortunately, such mechanisms adopt single-headed convolutional structures, fail to fuse fine-grained part information. Besides, convolution-based schemes can maintain local positional information but cannot effectively pre-serve relative positions between parts. This paper proposes a new CNN-Transformer hyper architecture called the Person Retrieval with Conv-Transformer (PRCT). We integrate the multi-head self-attention and positional embedding module, which are the core ingredients of non-convolutional Transformer, with a CNN-based part-feature extractor to maintain more precise within-part consistency in feature aggregation. With PRCT, we can effectively eliminate the part mis-alignments when matching different samples. We conduct extensive evaluations on the MSMT17, DukeMTMC-ReID, and Market-1501 datasets and obtain state-of-the-art performance.

Searching accuracy and efficiency are two challenges in person and vehicle Re-identification (Re-ID), which one focuses on robust representations learning which usually generating high-dimensional features while the other has not been fully explored. Hashing is a suitable solution to make REID efficient. However, directly extending the existing hashing methods to fast Re-ID faces two challenges: one is the non-overlap between training and testing set which need more discriminative hash codes, the other is the large identities in Re-ID tasks which will lead to slow convergence and hard optimization. In this work, we propose an attention pooling operator to exploit both local and global visual attributes which can break limited discriminative power in hash methods. To further make training procedure converge faster and optimize the network more easily, we substitute non-differentiable l1-regularization with smooth l1-regularization. In experiments, our work outperforms state-of-the-art hashing and quantization methods on both person and vehicle Re-ID datasets. Besides, the results can serve as a strong baseline in the field of deep hashing for fast Re-ID.11Code is available at https://github.com/cynthia951031/Hashing ReID.