TRAIL: Trajectory-Aware Visual Place Recognition against Unordered Databases
针对特征贫乏环境下视觉定位问题，提出TRAIL方法利用轨迹上下文信息，通过条件随机场结合视觉相似性和相机运动一致性提高定位准确性。
0 citationsRead paper
Selected work
Representative Papers
SAR Object Detection with Self-Supervised Pretraining and Curriculum-Aware Sampling
Small-object detection in spaceborne SAR imagery faces critical challenges including severe scarcity of annotated data, extreme scale imbalance, and poor model generalization. Method: This paper proposes TRANSAR, an end-to-end vision Transformer framework. It introduces (i) a novel mask image modeling (MIM)-based self-supervised pretraining paradigm leveraging over 25,700 km² of unlabeled SAR data; (ii) an auxiliary binary semantic segmentation head to improve small-object localization; and (iii) a curriculum-learning-driven dynamic class sampling scheduler to mitigate severe scale imbalance. Results: On multiple SAR benchmark datasets, TRANSAR consistently outperforms state-of-the-art models—including DeepLabv3, UNet, and SegFormer—with up to a 12.6% absolute gain in average precision (AP) for small objects. These results validate the effectiveness and advancement of the synergistic integration of self-supervised pretraining and curriculum-aware sampling.
0 citationsRead paper
Recent publications
Latest Papers
TRAIL: Trajectory-Aware Visual Place Recognition against Unordered Databases
针对特征贫乏环境下视觉定位问题，提出TRAIL方法利用轨迹上下文信息，通过条件随机场结合视觉相似性和相机运动一致性提高定位准确性。
0 citationsRead paper
SAR Object Detection with Self-Supervised Pretraining and Curriculum-Aware Sampling
Small-object detection in spaceborne SAR imagery faces critical challenges including severe scarcity of annotated data, extreme scale imbalance, and poor model generalization. Method: This paper proposes TRANSAR, an end-to-end vision Transformer framework. It introduces (i) a novel mask image modeling (MIM)-based self-supervised pretraining paradigm leveraging over 25,700 km² of unlabeled SAR data; (ii) an auxiliary binary semantic segmentation head to improve small-object localization; and (iii) a curriculum-learning-driven dynamic class sampling scheduler to mitigate severe scale imbalance. Results: On multiple SAR benchmark datasets, TRANSAR consistently outperforms state-of-the-art models—including DeepLabv3, UNet, and SegFormer—with up to a 12.6% absolute gain in average precision (AP) for small objects. These results validate the effectiveness and advancement of the synergistic integration of self-supervised pretraining and curriculum-aware sampling.
0 citationsRead paper