Yang Zheng

Nan Chen, Mengqi Huang, Zhuowei Chen et al.

Subject-driven text-to-image (T2I) customization has drawn significant interest in academia and industry. This task enables pre-trained models to generate novel images based on unique subjects. Existing studies adopt a self-reconstructive perspective, focusing on capturing all details of a single image, which will misconstrue the specific image's irrelevant attributes (e.g., view, pose, and background) as the subject intrinsic attributes. This misconstruction leads to both overfitting or underfitting of irrelevant and intrinsic attributes of the subject, i.e., these attributes are over-represented or under-represented simultaneously, causing a trade-off between similarity and controllability. In this study, we argue an ideal subject representation can be achieved by a cross-differential perspective, i.e., decoupling subject intrinsic attributes from irrelevant attributes via contrastive learning, which allows the model to focus more on intrinsic attributes through intra-consistency (features of the same subject are spatially closer) and inter-distinctiveness (features of different subjects have distinguished differences). Specifically, we propose CustomContrast, a novel framework, which includes a Multilevel Contrastive Learning (MCL) paradigm and a Multimodal Feature Injection (MFI) Encoder. The MCL paradigm is used to extract intrinsic features of subjects from high-level semantics to low-level appearance through crossmodal semantic contrastive learning and multiscale appearance contrastive learning. To facilitate contrastive learning, we introduce the MFI encoder to capture cross-modal representations. Extensive experiments show the effectiveness of CustomContrast in subject similarity and text controllability.

Jiachen Qian, Yang Zheng

We consider the problem of online prediction for an unknown, non-explosive linear stochastic system. With a known system model, the optimal predictor is the celebrated Kalman filter. In the case of unknown systems, existing approaches based on recursive least squares and its variants may suffer from degraded performance due to the highly imbalanced nature of the regression model. This imbalance can easily lead to overfitting and thus degrade prediction accuracy. We tackle this problem by injecting an inductive bias into the regression model via {exponential forgetting}. While exponential forgetting is a common wisdom in online learning, it is typically used for re-weighting data. In contrast, our approach focuses on balancing the regression model. This achieves a better trade-off between {regression} and {regularization errors}, and simultaneously reduces the {accumulation error}. With new proof techniques, we also provide a sharper logarithmic regret bound of $O(\log^3 N)$, where $N$ is the number of observations.

Yang Zheng, Izzat H. Izzat, Shahrzad Ziaee

Pedestrian detection is an essential task in autonomous driving research. In addition to typical color images, thermal images benefit the detection in dark environments. Hence, it is worthwhile to explore an integrated approach to take advantage of both color and thermal images simultaneously. In this paper, we propose a novel approach to fuse color and thermal sensors using deep neural networks (DNN). Current state-of-the-art DNN object detectors vary from two-stage to one-stage mechanisms. Two-stage detectors, like Faster-RCNN, achieve higher accuracy, while one-stage detectors such as Single Shot Detector (SSD) demonstrate faster performance. To balance the trade-off, especially in the consideration of autonomous driving applications, we investigate a fusion strategy to combine two SSDs on color and thermal inputs. Traditional fusion methods stack selected features from each channel and adjust their weights. In this paper, we propose two variations of novel Gated Fusion Units (GFU), that learn the combination of feature maps generated by the two SSD middle layers. Leveraging GFUs for the entire feature pyramid structure, we propose several mixed versions of both stack fusion and gated fusion. Experiments are conducted on the KAIST multispectral pedestrian detection dataset. Our Gated Fusion Double SSD (GFD-SSD) outperforms the stacked fusion and achieves the lowest miss rate in the benchmark, at an inference speed that is two times faster than Faster-RCNN based fusion networks.