Event detection is essential for surveillance, particularly in retail loss prevention, where accurate and timely monitoring is critical. Vision Language Models (VLMs) provide strong generalization but are inefficient at processing full video streams and are prone to hallucinations induced by redundant frames. We present SmartEyes, a plug-and-play system for real-time retail surveillance. SmartEyes introduces the Perception Cognition Focusing (PCF) framework, which combines lightweight perception with semantic triggering to isolate two keyframes (customer contact and departure) and constrains the VLMs to a focused differencing task. This design reduces hallucination by 44% compared to vanilla VLMs. From the demonstrated retail application, the proposed perception-to-reasoning pipeline is general and directly extends to industrial environments that require reliable event detection and real-time decision-making. Our demo includes a user-friendly Region of Interest (ROI) selection interface and live CCTV monitoring, producing accurate alerts within 1–2 seconds on a single RTX 4080 GPU. This lightweight framework design enables efficient deployment to broader industrial applications.

Accurately labeling (or annotation) data is still a bottleneck in computer vision, especially for large-scale tasks where manual labeling is time-consuming and error-prone. While tools like LabelImg can handle the labeling task, some of them still require annotators to manually label each image. In this paper, we introduce BakuFlow, a streamlining semi-automatic label generation tool. Key features include (1) a live adjustable magnifier for pixel-precise manual corrections, improving user experience; (2) an interactive data augmentation module to diversify training datasets; (3) label propagation for rapidly copying labeled objects between consecutive frames, greatly accelerating annotation of video data; and (4) an automatic labeling module powered by a modified YOLOE framework. Unlike the original YOLOE, our extension supports adding new object classes and any number of visual prompts per class during annotation, enabling flexible and scalable labeling for dynamic, real-world datasets. These innovations make BakuFlow especially effective for object detection and tracking, substantially reducing labeling workload and improving efficiency in practical computer vision and industrial scenarios.

In distributed manufacturing environments, Auto-mated Guided Vehicles (AGVs) equied with visual camera play a crucial role in automating material handling and optimizing production efficiency. Detecting anomalies during AGV operation is crucial to prevent potential malfunctions that could disrupt industrial processes. However, anomaly detection is challenging due to privacy concerns and the heterogeneity of data collected by AGVs across different factories. While sharing data across factories can improve the generalization capabilities of models, this can lead to overgeneralization in reconstruction-based anomaly detection, where the model reconstructs both normal and anomalous data too well, reducing its ability to detect anomalies. To address this problem, we propose FedCali, a federated learning framework that balances generalization and specialization across AGVs monitoring different manufacturing processes. Our proposed Gradient Guiding Mechanism (GGM) selectively aligns local model gradients with global knowledge only when necessary. This allows local models to retain their unique characteristics while benefiting from shared insights. Experiments with the MVTec dataset show that FedCali improves both reconstruction quality and anomaly detection accuracy, achieving higher AUROC scores and lower losses compared to baseline methods. This shows that FedCali is able to effectively process various manufacturing data collected by AGVs while maintaining data privacy.