Jun Sun

OpenDRIVE maps, widely used in autonomous driving simulation and validation, often suffer from missing intersections due to incomplete survey coverage or construction zones, significantly hindering downstream planning and verification tasks. We propose a novel large language model (LLM)-based framework, augmented with OpenDRIVE domain-specific geometric modification tools, to complete these missing intersections. Our zero-shot approach leverages the LLM as both a planner and a coder, sequentially invoking these tools to reconstruct road topology. We also introduce a data distillation-based fine-tuning method and a new point-alignment (centimeter-level) based evaluation methodology, optimizing performance with a lightweight LLM. To validate our methods, we further curate a balanced dataset comprising 137 pairs of OpenDRIVE maps from OpenStreetMap (OSM) covering diverse scenarios including urban intersections, highway ramps, and roundabouts. Experimental results demonstrate that our fine-tuned Qwen3-1.7B model achieved a 99.83% Completion Accuracy for completing missing intersections, generating high-precision OpenDRIVE data. This showcases our framework’s superior accuracy and efficiency in robustly completing OpenDRIVE maps.

With the prevalence of state-of-the-art generative models, photorealistic synthetic images can now be easily generated. However, the generated images may replicate contents from the original training images, which can lead to potential legal issues. In this paper, we propose a novel method called Conceptual-Similar-guided Self-supervised Language-Image Pre-training (CS-SLIP) that leverages both image and text modalities for the generated image provenance. Besides the self-supervised learning branch and contrastive learning branch, a conceptual-similar branch is designed to guide the model to learn a better feature representation of image-text-pairs. We also adopt the re-ranking method to refine the initial matching candidates via the cross-modal bi-directional retrieval. Extensive qualitative and quantitative experiments are conducted, which demonstrate that the replication indeed exists in the generated images, and our proposed method can effectively retrieve the most similar images from the training corpus to achieve the goal of generated image provenance analysis.

Recently, Deep Neural Networks (DNN) shows potential in many applications, while the risks behind DNN hinder its wide deployment. One of the most prominent risks of DNN is its fairness issue and it is necessary to mitigate the biases behind DNN to make it more trustworthy. Existing bias mitigation methods for learning fair feature representations trained from large-scale datasets through the correlation between the target attribute and sensitive attribute. However, these methods require many human annotations and ignore many latent biased attributes that are difficult to realize and discover by human means. To solve these problems, we propose a novel latent bias mitigation method, and the main idea is as follows: (1) amplification of a biased model without any sensitive attribute labels; (2) proposal of a Fair Contrastive Learning Loss to encourage learning of fair representations and discourage learning of biased information. Experimental results show that our method outperforms state-of-the-art benchmarks in term of latent bias mitigation, as well as achieving better accuracy.

Food-related issues have attracted increasing attention recently due to its various applications in our daily life, for example, restaurant service and dietary assessment, which are essential for human health management and understanding food characteristics. Many existing datasets are proposed for various food-relevant tasks, e.g., food detection and segmentation. However, most of the datasets are labelled with only dish-level annotations and lack detailed information for the corresponding ingredients. Hyperspectral imaging (HSI), which can explore the emissive characteristics of different objects with a long region of spectral bands, is of great potential for food and ingredient analysis. In the present work, a new food image dataset HSIFoodIngr-64 containing 3,389 pairs of HSI and RGB images with 21 dish classes and 64 ingredient categories was established. It is the first HSI-based food image dataset, of which all images are labelled with dish-level and ingredient-level annotations and corresponding pixel-wise ingredient masks. Therefore, our dataset can be applied to different food-centric tasks. Furthermore, this research proposed a benchmark method on the food ingredient retrieval task, which consists of two sub-networks called IPSN for ingredient prediction sub-network and ERSN for edge refinement sub-network. Extensive experiments were conducted on HSIFoodIngr-64 dataset and the effectiveness of our proposed method was verified. HSIFoodIngr-64 dataset is expected to provide a new perspective on food analysis and inspire more efforts on various food-related issues. We have made the HSIFoodIngr-64 dataset public at: https://doi.org/10.7910/DVN/E7WDNQ.

Automatic video understanding is a crucial piece of technology which promotes urban sustainability. Video grounding is a fundamental component of video understanding that has been evolving quickly in recent years, but its use is restricted due to the high labeling costs and typical performance limitations imposed by the pre-defined training dataset. In this paper, a novel atom-based zero-shot video grounding (AZVG) method is proposed to retrieve the segments in the video that correspond to a given input sentence. Although it is training-free, the performance of AZVG is competitive to the weakly supervised methods and better than unsupervised SOTA methods on the Charades-STA dataset. The method can support flexible queries as well as different video content. It can play an important role in a wider range of urban living applications.

Image segmentation is a long-standing problem in medical image analysis to facilitate the clinical diagnosis and intervention. Progress has been made owing to deep learning via supervised training with elaborate human labelling, however, the segmentation models trained by the labeled source domain cannot perform well in the target domain, making existing approaches lack robustness and generalization ability. Considering the acquisition of medical image labels is quite expensive and time-consuming, we propose a novel feature disentanglement-based unsupervised domain adaptation (UDA) method to improve the robustness of the trained model in the target domain. A segmentation network is designed to learn disentangled features with two parts: I. content-related features, which are responsible for the segmentation task and invariant across domains; II. style-related features, which elucidate the discrepancy between different domains. Feature disentanglement (FD) is achieved by multi-task learning and image translation. Meanwhile, knowledge distillation is introduced to improve the performance on fine-grained segmentations. And for objects with regular shape, we incorporate the adversarial training to predict shape-invariant segmentation masks across domains. Comprehensive experiments are conducted on retina vessel segmentation and sinus surgical instrument segmentation to validate the effectiveness of the proposed method. The average Dice of twenty regular transfer directions achieves 79.26% on five public benchmarks of retina vessel segmentation, the average Dice of two transfer directions from regular to UWF attains 72.63%, and the Dice from cadaveric images to live images reaches 68.1% on sinus surgical instrument segmentation. The results demonstrate that the proposed method achieves the state-of-the-art segmentation performance in the UDA setting.