Search

Small object detection is a fundamental and challenging issue in computer vision. We believe that there are two factors that affect the performance of small object detection: small object dataset and small object itself. In terms of datasets, we introduce a dataset named SeaDefine, which opens up a new direction for small object detection in maritime environment. For the small object itself, we utilize deep feature learning and feature fusion network (DFLFFN) to help detect objects. Concretely, the designed deep feature learning module (DFLM) at the single-layer level can describe objects for a variety of scenarios through activating multi-scale receptive fields over a wider scope. Meanwhile, to intensify classification capacity of small objects, the shallow features with rich details will be integrated with the deep features generated from the DFLM by introducing feature fusion block (FFB). In addition, we analyze the multi-scale strategy from a mathematical perspective to a certain extent. A large number of results in the experiments show that proposed DFLFFN achieves the leading detection performance on the MS-COCO and SeaDefine datasets. In particular, our DFLFFN surpasses the baseline by 4.1 points on APrs score for SeaDefine dataset, and 7.8 points on AP S score for MS-COCO dataset. [ABSTRACT FROM AUTHOR]

Small object detection is a challenging task in computer vision. We claim that the huge performance gap between the small object detectors and normal sized object detectors stems from two aspects, including the small object dataset and the small object itself. In terms of datasets, we build a large-scale dataset with high image resolution dubbed Small-PCB, in order to promote detection in semiconductor industry. For the small object itself, we utilize multi-scale feature learning and feature fusion strategy to help detect objects. More concretely, we devise two novel components to predict small objects better: re-parameterized module with channel shuffle (RMCS) and multi-scale feature enhanced convolution (MFEC). MFEC aims to split input channels into several parts and applies convolutions with different sizes to each part, and adopt point-by-point convolution to fuse individual channel features. RMCS not only use structural re-parameterization, but also channel shuffle. The usage of channel shuffle can be seen as a fusion of channel features. It strengthens feature information interaction between different channel groups, which bring more informative feature clues. Based on the RMCS and the MFEC, we introduce OIU-RMCS and M-MFEC, respectively. Finally, we build our I-YOLO via integrating these two components into a YOLO-based detector. A large number of qualitative and quantitative results in the experiments indicate that our proposed I-YOLO achieves the state-of-the-art performance on the popular AI-TODv2 and Small-PCB datasets. [ABSTRACT FROM AUTHOR]