Yiqun Duan

Yi Li, Hualiang Wang, Yiqun Duan et al.

Contrastive language-image pre-training (CLIP) is a powerful vision-language model that has shown great benefits for various tasks. However, we have identified some issues with its explainability, which undermine its credibility and limit the capacity for related tasks. Specifically, we find that CLIP tends to focus on background regions rather than foregrounds, with noisy activations at irrelevant positions on the visualization results. These phenomena conflict with conventional explainability methods based on the class attention map (CAM), where the raw model can highlight the local foreground regions using global supervision without alignment. To address these problems, we take a closer look at its architecture and features. Based on thorough analyses, we find the raw self-attentions link to inconsistent semantic regions, resulting in the opposite visualization. Besides, the noisy activations are owing to redundant features among categories. Building on these insights, we propose the CLIP Surgery for reliable CAM, a method that allows surgery-like modifications to the inference architecture and features, without further fine-tuning as classical CAM methods. This approach significantly improves the explainability of CLIP, surpassing existing methods by large margins. Besides, it enables multimodal visualization and extends the capacity of raw CLIP on open-vocabulary tasks without extra alignment. The code is available at https://github.com/xmed-lab/CLIP_Surgery.

Yiqun Duan, Xianda Guo, Zheng Zhu

Monocular depth estimation is a challenging task that predicts the pixel-wise depth from a single 2D image. Current methods typically model this problem as a regression or classification task. We propose DiffusionDepth, a new approach that reformulates monocular depth estimation as a denoising diffusion process. It learns an iterative denoising process to `denoise' random depth distribution into a depth map with the guidance of monocular visual conditions. The process is performed in the latent space encoded by a dedicated depth encoder and decoder. Instead of diffusing ground truth (GT) depth, the model learns to reverse the process of diffusing the refined depth of itself into random depth distribution. This self-diffusion formulation overcomes the difficulty of applying generative models to sparse GT depth scenarios. The proposed approach benefits this task by refining depth estimation step by step, which is superior for generating accurate and highly detailed depth maps. Experimental results on KITTI and NYU-Depth-V2 datasets suggest that a simple yet efficient diffusion approach could reach state-of-the-art performance in both indoor and outdoor scenarios with acceptable inference time.

Yiqun Duan, Zhen Wang, Yi Li et al.

Electroencephalograms (EEGs) are brain dynamics measured outside the brain, which have been widely utilized in non-invasive brain-computer interface applications. Recently, various neural network approaches have been proposed to improve the accuracy of EEG signal recognition. However, these approaches severely rely on manually designed network structures for different tasks which generally are not sharing the same empirical design cross-task-wise. In this paper, we propose a cross-task neural architecture search (CTNAS-EEG) framework for EEG signal recognition, which can automatically design the network structure across tasks and improve the recognition accuracy of EEG signals. Specifically, a compatible search space for cross-task searching and an efficient constrained searching method is proposed to overcome challenges brought by EEG signals. By unifying structure search on different EEG tasks, this work is the first to explore and analyze the searched structure difference cross-task-wise. Moreover, by introducing architecture search, this work is the first to analyze model performance by customizing model structure for each human subject. Detailed experimental results suggest that the proposed CTNAS-EEG could reach state-of-the-art performance on different EEG tasks, such as Motor Imagery (MI) and Emotion recognition. Extensive experiments and detailed analysis are provided as a good reference for follow-up researchers.

Yi Li, Hualiang Wang, Yiqun Duan et al.

Contrastive Language-Image Pre-training (CLIP) learns rich representations via readily available supervision of natural language. It improves the performance of downstream vision tasks, including but not limited to the zero-shot, long tail, segmentation, retrieval, caption, and video. However, the visual explainability of CLIP is rarely studied, especially for the raw feature map. To provide visual explanations of its predictions, we propose the Image-Text Similarity Map (ITSM). Based on it, we surprisingly find that CLIP prefers the background regions than the foregrounds, and shows erroneous visualization results against human understanding. This phenomenon is universal for both vision transformers and convolutional networks, which suggests this problem is unique and not owing to certain network. Experimentally, we find the devil is in the pooling part, where inappropriate pooling methods lead to a phenomenon called semantic shift. For this problem, we propose the Explainable Contrastive Language-Image Pre-training (ECLIP), which corrects the explainability via the Masked Max Pooling. Specifically, to avoid the semantic shift, we replace the original attention pooling by max pooling to focus on the confident foreground, with guidance from free attention during training. Experiments on three datasets suggest that ECLIP greatly improves the explainability of CLIP, and beyond previous explainability methods at large margins. The code will be released later.

Xianda Guo, Ruijun Zhang, Yiqun Duan et al.

Accurate spatial reasoning in outdoor environments - covering geometry, object pose, and inter-object relationships - is fundamental to downstream tasks such as mapping, motion forecasting, and high-level planning in autonomous driving. We introduce SURDS, a large-scale benchmark designed to systematically evaluate the spatial reasoning capabilities of vision language models (VLMs). Built on the nuScenes dataset, SURDS comprises 41,080 vision-question-answer training instances and 9,250 evaluation samples, spanning six spatial categories: orientation, depth estimation, pixel-level localization, pairwise distance, lateral ordering, and front-behind relations. We benchmark leading general-purpose VLMs, including GPT, Gemini, and Qwen, revealing persistent limitations in fine-grained spatial understanding. To address these deficiencies, we go beyond static evaluation and explore whether alignment techniques can improve spatial reasoning performance. Specifically, we propose a reinforcement learning-based alignment scheme leveraging spatially grounded reward signals - capturing both perception-level accuracy (location) and reasoning consistency (logic). We further incorporate final-answer correctness and output-format rewards to guide fine-grained policy adaptation. Our GRPO-aligned variant achieves an overall score of 40.80 in the SURDS benchmark. Notably, it outperforms proprietary systems such as GPT-4o (13.30) and Gemini-2.0-flash (35.71). To our best knowledge, this is the first study to demonstrate that reinforcement learning-based alignment can significantly and consistently enhance the spatial reasoning capabilities of VLMs in real-world driving contexts. We release the SURDS benchmark, evaluation toolkit, and GRPO alignment code through: https://github.com/XiandaGuo/Drive-MLLM.

Jinzhao Zhou, Yiqun Duan, Zhihong Chen et al.

Making sense of multiple modalities can yield a more comprehensive description of real-world phenomena. However, learning the co-representation of diverse modalities is still a long-standing endeavor in emerging machine learning applications and research. Previous generative approaches for multimodal input approximate a joint-modality posterior by uni-modality posteriors as product-of-experts (PoE) or mixture-of-experts (MoE). We argue that these approximations lead to a defective bound for the optimization process and loss of semantic connection among modalities. This paper presents a novel variational method on sets called the Set Multimodal VAE (SMVAE) for learning a multimodal latent space while handling the missing modality problem. By modeling the joint-modality posterior distribution directly, the proposed SMVAE learns to exchange information between multiple modalities and compensate for the drawbacks caused by factorization. In public datasets of various domains, the experimental results demonstrate that the proposed method is applicable to order-agnostic cross-modal generation while achieving outstanding performance compared to the state-of-the-art multimodal methods. The source code for our method is available online https://anonymous.4open.science/r/SMVAE-9B3C/.

Yi Li, Yiqun Duan, Zhanghui Kuang et al.

Weakly-Supervised Semantic Segmentation (WSSS) segments objects without a heavy burden of dense annotation. While as a price, generated pseudo-masks exist obvious noisy pixels, which result in sub-optimal segmentation models trained over these pseudo-masks. But rare studies notice or work on this problem, even these noisy pixels are inevitable after their improvements on pseudo-mask. So we try to improve WSSS in the aspect of noise mitigation. And we observe that many noisy pixels are of high confidence, especially when the response range is too wide or narrow, presenting an uncertain status. Thus, in this paper, we simulate noisy variations of response by scaling the prediction map multiple times for uncertainty estimation. The uncertainty is then used to weight the segmentation loss to mitigate noisy supervision signals. We call this method URN, abbreviated from Uncertainty estimation via Response scaling for Noise mitigation. Experiments validate the benefits of URN, and our method achieves state-of-the-art results at 71.2% and 41.5% on PASCAL VOC 2012 and MS COCO 2014 respectively, without extra models like saliency detection. Code is available at https://github.com/XMed-Lab/URN.

Xianda Guo, Ruijun Zhang, Yiqun Duan et al.

Depth estimation is a fundamental task for 3D scene understanding in autonomous driving, robotics, and augmented reality. Existing depth datasets, such as KITTI, nuScenes, and DDAD, have advanced the field but suffer from limitations in diversity and scalability. As benchmark performance on these datasets approaches saturation, there is an increasing need for a new generation of large-scale, diverse, and cost-efficient datasets to support the era of foundation models and multi-modal learning. We present ROVR, a large-scale, diverse, and cost-efficient depth dataset designed to capture the complexity of real-world driving. ROVR comprises 200K high-resolution frames across highway, rural, and urban scenarios, spanning day/night and adverse weather conditions. A lightweight acquisition pipeline ensures scalable collection, while sparse but statistically sufficient ground truth supports robust training. Benchmarking with state-of-the-art monocular depth models reveals severe cross-dataset generalization failures: models achieving near-ceiling accuracy on KITTI degrade drastically on ROVR, and even when trained on ROVR, current methods fall short of saturation. These results highlight the unique challenges posed by ROVR-scene diversity, dynamic environments, and sparse ground truth, establishing it as a demanding new platform for advancing depth estimation and building models with stronger real-world robustness. Extensive ablation studies provide a more intuitive understanding of our dataset across different scenarios, lighting conditions, and generalized ability.