Contemporary deep learning approaches for optical flow estimation continue to face persistent challenges in model interpretability, generalization capacity, and deployment efficiency, significantly constraining their practical implementation. This limitation becomes particularly critical in applications such as visual odometry (VO), where precise sparse point tracking supersedes the conventional emphasis on dense optical flow accuracy. Moreover, the lack of a joint framework combining keypoint detection and optical flow estimation limits sparse optical flow performance. To address these fundamental issues, we propose a novel dual-task imperative learning framework that synergistically optimizes sparse optical flow estimation (iFLOW) with adaptive keypoint detection (iPOINT). Our methodology implements an Expectation-Maximization (EM) paradigm where iFLOW and iPOINT undergo alternating optimization through a Gauss-Newton reasoning engine. This innovative architecture leverages convolutional feature advantages under the generalized feature invariance principle. The resulting imperative learning mechanism imbues our framework with enchanced interpretability and cross-domain adaptability while maintaining computational efficiency. Through comparative evaluations against classical and learning-based baselines, our ultra-compact models (0.05M parameters for iFLOW, 0.09M for iPOINT) demonstrate remarkable performance across multiple metrics (End-point Error, F1-all, VO trajectory accuracy) despite requiring only 200 training image pairs.

Existing continual learning works explored strategies like memory replay, regularization, and parameter isolation, but little analysis was conducted on the optimization behavior of LLMs’ continual fine-tuning. In this work, we investigate the geometric connections of different minima along the continual LLM fine-tuning trajectories. We validate this phenomenon on LLMs and propose a new method called Interpolation-based LoRA (I-LoRA). I-LoRA can strike a balance between plasticity and stability through parameter interpolation, which constructs a dual-memory experience replay framework based on LoRA. Experiments on eight domain-specific benchmarks demonstrate that I-LoRA consistently shows significant improvement over previous approaches with up to 11% performance gains. Our code is available at https://anonymous.4open.science/r/LLMCL-3823.

Open-vocabulary part segmentation (OVPS) struggles with structurally connected boundaries due to the inherent conflict between continuous image features and discrete classification mechanism. To address this, we propose PBAPS, a novel training-free framework specifically designed for OVPS. PBAPS leverages structural knowledge of object-part relationships to guide a progressive segmentation from objects to fine-grained parts. To further improve accuracy at challenging boundaries, we introduce a Boundary-Aware Refinement (BAR) module that identifies ambiguous boundary regions by quantifying classification uncertainty, enhances the discriminative features of these ambiguous regions using high-confidence context, and adaptively refines part prototypes to better align with the specific image. Experiments on Pascal-Part-116, ADE20K-Part-234, PartImageNet demonstrate that PBAPS significantly outperforms state-of-the-art methods, achieving 46.35\% mIoU and 34.46\% bIoU on Pascal-Part-116. Our code is available at https://github.com/TJU-IDVLab/PBAPS.

With the rapid advancement of technologies such as artificial intelligence and the Internet of Things, video surveillance—being a core component of video IoT systems—has been widely adopted for underground coal mine safety monitoring. However, the dim lighting and heavy coal dust in underground mines result in poor visibility and significant detail loss in monitoring images, posing a major challenge to coal mine safety management. To address these issues, we propose a low-light image enhancement method tailored for underground coal mine environments, based on Zero-DCE. In our method, traditional convolutions are replaced with Ghost modules to reduce computational cost while maintaining feature extraction capability. Additionally, we incorporate global context blocks and a Vision Transformer branch to integrate more global information into the model. Specifically, the global context blocks improve the model’s ability to correct uneven illumination and prevent overexposure. Meanwhile, the Vision Transformer branch captures long-range dependencies and fuses local and global features to enhance brightness while mitigating color distortion. Furthermore, we replace the original quadratic iterative function with a reciprocal illumination mapping function, enabling more stable and perceptually aligned brightness adjustments. Experimental results on the coal mine underground personnel dataset demonstrate that our method outperforms several state-of-the-art low-light enhancement techniques, achieving superior results in both qualitative and quantitative evaluations. These findings indicate that our approach significantly improves the visibility and overall quality of underground coal mine monitoring images.