Two-Stage Adaptive Network for Semi-Supervised Cross-Domain Crater Detection under Varying Scenario Distributions

TL;DR

This work tackles cross-domain crater detection under varying scenario distributions by introducing TAN, a two-stage adaptive network based on YOLOv5. Stage one combines ASAF with NAM to robustly fuse shallow and deep features and mitigate scale variance, while SHEM reduces overfitting to hard source-domain examples. Stage two employs SPF to generate high-quality pseudo-labels on the target domain and fine-tune the model, aided by a domain-aware augmentation strategy. Across DACD, LROC, and DOTA datasets, TAN demonstrates strong domain adaptation and generalisation, outperforming several state-of-the-art detectors in cross-domain crater detection and showing competitive performance on remote-sensing benchmarks. These results suggest significant practical potential for robust crater discovery on unseen planetary surfaces with limited labeled data.

Abstract

Crater detection can provide valuable information for humans to explore the topography and understand the history of extraterrestrial planets. Due to the significantly varying scenario distributions, existing detection models trained on known labelled crater datasets are hardly effective when applied to new unlabelled planets. To address this issue, we propose a two-stage adaptive network (TAN) for semi-supervised cross-domain crater detection. Our network is built on the YOLOv5 detector, where a series of strategies are employed to enhance its cross-domain generalisation ability. In the first stage, we propose an attention-based scale-adaptive fusion (ASAF) strategy to handle objects with significant scale variances. Furthermore, we propose a smoothing hard example mining (SHEM) loss function to address the issue of overfitting on hard examples. In the second stage, we propose a sort-based pseudo-labelling fine-tuning (SPF) strategy for semi-supervised learning to mitigate the distributional differences between source and target domains. For both stages, we employ weak or strong image augmentation to suit different cross-domain tasks. Experimental results on benchmark datasets demonstrate that the proposed network can enhance domain adaptation ability for crater detection under varying scenario distributions.

Figures (8)

• Figure S1: (Top) One of the samples in the LROC dataset and the distributions of all craters in this dataset. (Bottom) One of the samples in the DACD dataset and the distributions of all craters in this dataset. Compared with the top sample, the bottom one has more background interference. According to the statistical results of these two datasets, the LROC dataset contains smaller and more craters than DACD.
• Figure S2: The architecture of our proposed two-stage TAN model for cross-domain crater detection. The first stage: (1) ASAF utilises the attention-based NAM (Normalisation-Based Attention Module) to fuse shallow information to improve scale adaptation abilities; (2) the SHEM loss function is used to alleviate the bias of the model. The second stage: The SPF strategy is adopted to sort and select high-quality pseudo-labels which are used to fine-tune the model. In these two stages, we adopt weak or strong image augmentation to match different cross-domain tasks. The new components are highlighted in blue font or regions.
• Figure S3: Illustration of our model architecture with the ASAF strategy. To prevent the model from losing crucial information, we incorporate C3TR (C3 + Transformer) into the backbone. We pass the shallow feature maps of different stages to NAM to obtain shallow attention-based feature maps. We also inject multiple C3 modules into the neck to detect large targets.
• Figure S4: The architecture of NAM nam.
• Figure S5: The overall flow of the SHEM loss function. We calculate BFLs for the feature maps at four scales, each with a distinct distribution of loss values. Then, we select the top K% of loss values that have been sorted for the feature maps. Subsequently, we average and weigh the loss values at different scales to obtain the Loss Rank Function (LRM), followed by L2 regularisation to obtain the objectness loss.