Benefiting from advancements in generic object detectors, significant progress has been achieved in the field of face detection. Among these algorithms, the You Only Look Once (YOLO) series plays an important role due to its low training computation cost. However, we have observed that face detectors based on lightweight YOLO models struggle with accurately detecting small faces. This is because they preserve more semantic information for large faces while compromising the detailed information for small faces. To address this issue, this study makes two contributions to enhance detection performance, particularly for small faces: (1) modifying the neck part of the architecture by integrating a Gather-and-Distribute mechanism instead of the traditional Feature Pyramid Network to tackle the information fusion challenges inherent in YOLO-based models; and (2) incorporating an additional detection head specifically designed for detecting small faces. To evaluate the performance of the proposed face detector, we introduce a new dataset named XD-Face for the face detection task. In the experimental section, the proposed model is trained using the Wider Face dataset and evaluated on both Wider Face and XD-face datasets. Experimental results demonstrate that the proposed face detector outperforms other excellent face detectors across all datasets involving small faces and achieved improvements of 1.1%, 1.09%, and 1.35% in the AP50 metric on the WiderFace validation dataset compared to the baseline YOLOv5s-based face detector.
Place recognition refers to the problem of finding the position of a query image based on a series of images acquired at different places. Yet the day and night place recognition problem is hard to solve due to the illumination and appearance changes. Image-to-image translation methods have been introduced to solve the place recognition problem by synthesizing daytime images from the night ones. However, these methods cannot achieve good translation performance with low-quality night-time images. In this paper, a new method is introduced to improve the quality of night-time restored images by combining image enhancement and image inpainting methods. Three kinds of enhanced night-time images are generated based on the proposed method. Our place recognition system includes a model of GoogleNet to generate deep features of input images and nearest neighbor searching for the image retrieval process. The approach is tested on the Oxford RobotCar dataset, where three low-quality night sequences are selected as query sequences, and a day sequence is selected as a reference sequence. The results obtained with the approach based on the three proposed enhanced night-time images are better than those obtained with the raw night-time images. The results of our proposed place recognition system are also compared with two state-of-art place recognition methods: ToDayGAN and densevlad.
To address face occlusion, low detection rates of small-scale faces, and sample imbalance in dense visual scenarios, we propose a YOLOv7-based detector with four key improvements: (1) an optimized MPConv module to enhance feature extraction; (2) a novel CFPM to boost sensitivity to occluded samples; (3) an integration of the DyHead block in IDetect to mitigate feature loss from sample imbalance; (4) an SW-SCE loss function with a dual-input network to better detect small faces. Experiments on the WiderFace dataset show that our method improves detection performance by 1.2%, 1.8%, and 3% on the easy, medium, and hard subsets over the baseline. These gains strengthen face detection in dense, challenging environments with heavy occlusion and small-scale targets.
Existing video action recognition methods face two challenges when applied to classroom surveillance: first, they struggle to balance local spatial and global temporal information; second, they are affected by imbalanced category distribution where routine behaviors dominate while educationally significant rare behaviors are often missed. To address these issues, this paper proposes two components. A Regional Global-Time Featurizer (RGTF) extracts compact local spatial and global temporal features via frequency-domain transformation with low computational overhead. An Exponential Focal Loss (EFL) adaptively reweights hard samples to mitigate the impact of imbalanced data. Experiments on XDCR, AVA2.2, Kinetics-400, UCF101, and Something-Something V2 show that RGTF improves baseline models by up to 1.8% and EFL outperforms standard losses by up to 0.92%. An optional LLM-assisted module is additionally provided as an application example to illustrate one possible way of using the recognition outputs for generating qualitative classroom feedback. This module is independent of the core recognition pipeline.