[Paper Review] Progressive Domain Adaptation with Contrastive Learning for Object Detection in the Satellite Imagery

Debojyoti Biswas, Jelena Tešić|arXiv (Cornell University)|Sep 6, 2022

Advanced Neural Network Applications4 citations

TL;DR

This paper proposes LGDA, a progressive domain adaptation framework with contrastive learning for small object detection in satellite imagery. By integrating spatial pyramid pooling, heatmap-based region proposal, and a novel difficulty-aware focal loss, it improves feature learning and domain alignment through local and global contrastive learning, achieving up to 7.4% higher mAP than state-of-the-art methods on DOTA and NWPU VHR-10 benchmarks.

ABSTRACT

State-of-the-art object detection methods applied to satellite and drone imagery largely fail to identify small and dense objects. One reason is the high variability of content in the overhead imagery due to the terrestrial region captured and the high variability of acquisition conditions. Another reason is that the number and size of objects in aerial imagery are very different than in the consumer data. In this work, we propose a small object detection pipeline that improves the feature extraction process by spatial pyramid pooling, cross-stage partial networks, heatmap-based region proposal network, and object localization and identification through a novel image difficulty score that adapts the overall focal loss measure based on the image difficulty. Next, we propose novel contrastive learning with progressive domain adaptation to produce domain-invariant features across aerial datasets using local and global components. We show we can alleviate the degradation of object identification in previously unseen datasets. We create a first-ever domain adaptation benchmark using contrastive learning for the object detection task in highly imbalanced satellite datasets with significant domain gaps and dominant small objects. The proposed method results in a 7.4% increase in mAP performance measure over the best state-of-art.

Motivation & Objective

• Address the challenge of detecting small and densely packed objects in highly variable satellite and drone imagery.
• Overcome domain shift between source and target satellite datasets due to geographical, lighting, and acquisition condition variations.
• Improve feature representation and model generalization in low-resource, imbalanced aerial datasets using contrastive learning.
• Develop a domain adaptation pipeline that progressively aligns features across multiple intermediate domains to reduce domain gap.
• Achieve state-of-the-art performance on benchmark satellite datasets with significant domain gaps and class imbalance.

Proposed method

• Introduce a heatmap-based region proposal network (RPN) to better localize small and dense objects in aerial images.
• Design a difficulty-aware focal loss that dynamically adjusts loss weighting based on image-level difficulty to handle hard examples.
• Implement local feature alignment via contrastive learning (LDA) to align fine-grained, spatially localized features between source and target domains.
• Propose global domain alignment (GDA) using contrastive learning on high-receptive-field features to capture semantic-level domain invariance.
• Integrate both LDA and GDA modules into a unified progressive domain adaptation framework (LGDA) with multi-scale feature fusion.
• Optimize hyperparameters including negative sample count (7), temperature value (12% penalty), and training schedule (60 epochs) for robust contrastive learning.

Experimental results

Research questions

• RQ1 Can contrastive learning effectively reduce domain shift in satellite imagery with high geographical and environmental variability?
• RQ2 How does progressive domain adaptation with intermediate domains improve generalization in small object detection across diverse aerial datasets?
• RQ3 To what extent does difficulty-aware loss reweighting enhance detection performance on imbalanced, high-variability satellite datasets?
• RQ4 What hyperparameter settings (e.g., negative sample count, temperature) yield optimal contrastive learning performance in small object detection?
• RQ5 Does combining local and global contrastive alignment lead to better mAP than single-level alignment in satellite object detection?

Key findings

• The proposed LGDA method achieves 46.7% mAP on the DOTA target dataset, outperforming the best existing method (MGADA) by +7.3% mAP.
• On the NWPU VHR-10 dataset, LGDA achieves 71.4% mAP, representing a +4.6% improvement over MGADA.
• Local contrastive learning (LDA) with a weight of 0.1 on the loss function yields over 8% and 12% mAP gains on DOTA and NWPU, respectively.
• Global domain alignment (GDA) with a lower weight (0.01) outperforms LDA, indicating the importance of high-level semantic alignment.
• Using 7 negative samples per contrastive loss achieves optimal balance; increasing beyond this harms performance due to false negatives in imbalanced data.
• Training for 60 epochs with a 12% penalty on contrastive loss yields the best performance, while further training or higher penalties degrade results.