Abstract:The ambiguous appearance, tiny scale, and fine-grained classes of objects in remote sensing imagery inevitably lead to the noisy annotations in category labels of detection dataset. However, the effects and treatments of the label noises are underexplored in modern oriented remote sensing object detectors. To address this issue, we propose a robust oriented remote sensing object detection method through dynamic loss decay (DLD) mechanism, inspired by the two phase ``early-learning'' and ``memorization'' learning dynamics of deep neural networks on clean and noisy samples. To be specific, we first observe the end point of early learning phase termed as EL, after which the models begin to memorize the false labels that significantly degrade the detection accuracy. Secondly, under the guidance of the training indicator, the losses of each sample are ranked in descending order, and we adaptively decay the losses of the top K largest ones (bad samples) in the following epochs. Because these large losses are of high confidence to be calculated with wrong labels. Experimental results show that the method achieves excellent noise resistance performance tested on multiple public datasets such as HRSC2016 and DOTA-v1.0/v2.0 with synthetic category label noise. Our solution also has won the 2st place in the "fine-grained object detection based on sub-meter remote sensing imagery" track with noisy labels of 2023 National Big Data and Computing Intelligence Challenge.

What problem does this paper attempt to address?

The paper primarily addresses the problem of object detection in remote sensing images, particularly proposing a solution in the presence of class label noise. Specifically, the paper tackles the following core issues: 1. **Background and Challenges**: When performing object detection in remote sensing images, the issues of object appearance ambiguity, small object sizes, and fine-grained class divisions lead to prevalent label noise in datasets. These problems significantly reduce the detection accuracy of models. 2. **Limitations of Existing Work**: Most existing oriented remote sensing object detection (ORSOD) methods focus on improving detection frameworks and network components, with few methods specifically addressing the prevalent issue of class label noise in datasets. 3. **Proposed Solution**: To address the above issues, the paper proposes a robust oriented remote sensing object detection method through a Dynamic Loss Decay (DLD) mechanism. This method is inspired by the learning dynamics of deep neural networks (DNNs) in the "early learning" and "memorization" phases on clean and noisy samples. 4. **Key Steps**: - **Early Learning Phase Identification**: First, the paper identifies the endpoint of the "early learning" phase (EL), where the model starts memorizing incorrect labels, by monitoring the accuracy curve changes during training. - **Dynamic Loss Decay**: Second, during the "memorization" phase, the loss weights of the top K samples with the highest loss values (which are likely to have incorrect labels) are gradually reduced in each round to minimize their impact on the model. 5. **Experimental Validation**: The effectiveness and robustness of the proposed method are validated through experiments on multiple public datasets, including HRSC2016 and DOTA-v1.0/v2.0 with synthetic class label noise. Additionally, the method achieved second place in the relevant track of the 2023 National Big Data and Computing Intelligence Challenge. In summary, this paper aims to address the issue of detection accuracy degradation caused by class label noise in remote sensing images and proposes an effective solution—a robust oriented remote sensing object detection method based on the Dynamic Loss Decay mechanism.

Dynamic Loss Decay based Robust Oriented Object Detection on Remote Sensing Images with Noisy Labels

Robust Tiny Object Detection in Aerial Images amidst Label Noise

Remote Sensing Object Detection in Disturbed Environment Based on Improved RTMDet Network

Feature Reassembly and Self-Attention for Oriented Object Detection in Remote Sensing Images

Consistency-Regularized Learning for Remote Sensing Scene Classification with Noisy Labels

Dynamic Sensing and Correlation Loss Detector for Small Object Detection in Remote Sensing Images

MOL: Towards Accurate Weakly Supervised Remote Sensing Object Detection Via Multi-view Noisy Learning

A Robust Training Method for Object Detectors in Remote Sensing Image

A Self-Training Method with a Noise-Robust Loss Function for Weakly Supervised Object Detection

RSDS: A Specialized Loss Calculation Method for Dense Small Object Detection in Remote Sensing Images

Towards Noise-resistant Object Detection with Noisy Annotations

Noise-Tolerant Deep Neighborhood Embedding for Remotely Sensed Images with Label Noise

Learning Adversarially Robust Object Detector with Consistency Regularization in Remote Sensing Images

Guiding Clean Features for Object Detection in Remote Sensing Images

Hierarchical Mask Prompting and Robust Integrated Regression for Oriented Object Detection

Mining Oriented Information for Semi-Supervised Object Detection in Remote Sensing Images

GSDDet: Ground Sample Distance Guided Object Detection for Remote Sensing Images

Exploiting Label Uncertainty for Enhanced 3D Object Detection From Point Clouds

Deep Active Learning for Remote Sensing Object Detection

Oriented Object Detection in Remote Sensing Image Based on YOLOV5

OD-YOLO: Robust Small Object Detection Model in Remote Sensing Image with a Novel Multi-Scale Feature Fusion