Instance Relation Graph Guided Source-Free Domain Adaptive Object Detection

• Source-free adaptation addresses settings where privacy, transmission, or proprietary concerns restrict source-data access during adaptation.The introduction contrasts transmitting a source-trained detector model (∼500-1000MB) with source data (∼10-100GB).Source: §1 · p. 1
• Noisy source-model pseudo-labels and the absence of source data make adaptation challenging, motivating target-information distillation and stronger target representations.Source: §1 · pp. 1–2
• For a FoggyCityscapes bus, proposals with high overlap to the ground truth predict car, truck, or bus inconsistently, illustrating feature sensitivity to small proposal shifts.Source: §1 · p. 2
• RPN proposals provide multiple views of an object for contrastive learning, but their class-agnostic predictions make positive and negative pair assignment difficult.The paper notes that detection-oriented contrastive learning also faces high GPU and memory costs from large batches and multiple views.Source: Figure 3 · p. 2, §1 · p. 2
• A graph convolution network models relations among RPN proposals and uses them to identify positive and negative pairs that guide contrastive learning.Each graph node corresponds to a proposal, and its edges encode proposal similarity relations.Source: §1 · p. 2
• The method is evaluated on multiple visually distinct detection benchmarks and outperforms existing source-free and many unsupervised domain adaptation methods.Source: §1 · p. 2

Unsupervised Domain Adaptation

• Object-detection UDA methods use feature alignment, pseudo-label self-training, image generation, or object-center-aware training, while assuming access to labeled source and unlabeled target data.Source: §2 · p. 3

Source-Free Domain Adaptation

• Prior SFDA work spans classification, segmentation, detection, and video segmentation; detection approaches include pseudo-label self-training and self-supervised feature learning.Source: §2 · p. 3

Contrastive Learning

• Contrastive learning pulls positive-pair representations together and separates negatives, but detection extensions are difficult because models are computationally expensive and large batches are costly.Source: §2 · p. 3

Graph Convolution Neural Networks

• IoU-based graphs can miss relations between nonoverlapping proposals, such as two sidecar proposals that should be treated as a positive pair for contrastive learning.The paper motivates a learnable graph convolution network to model inter-instance relations instead.Source: §2 · p. 3

3.1. Preliminaries

• SFDA uses a source-trained model and unlabeled target data; mean-teacher self-training supervises the student with filtered teacher pseudo-labels, which remain noisy under domain shift.The teacher is progressively updated by exponential moving average of student weights.Source: §3.1 · pp. 3–4

3.2. Graph-guided contrastive learning

• RPN proposals provide multiple views of target instances, while IRG models their relations to identify positive and negative proposal pairs for contrastive learning.Without target ground-truth boxes, class-agnostic RPN predictions do not reveal which proposals belong to the same instance.Source: §3.2 · pp. 4–5

3.2.1 Instance Relation Graph (IRG)

• IRG represents RPN proposal RoI features as graph nodes and learns interproposal relations; Graph Distillation Loss aligns class logits across graph-transformed and original features and student-teacher pipelines.Both networks share IRG, which is constructed from teacher proposals.Source: §3.2.1 · p. 5

3.2.2 Graph Contrastive Loss (GCL)

• Graph Contrastive Loss uses thresholded IRG relations to label proposal pairs, then combines those labels with key-query pairwise logits to train the student on target features.The teacher parameters are updated via EMA, while GCL updates only the student.Source: §3.2.2 · pp. 5–6

3.3. Overall objective

• The overall SFDA objective combines student pseudo-label loss, graph distillation loss, and graph contrastive loss.These terms are denoted SL, LGDL, and LGCL, respectively.Source: §3.3 · p. 6

4. Experiments and Results

• Experiments compare UDA and SFDA methods across four shifts: adverse weather, real-to-artistic, synthetic-to-real, and cross-camera.UDA has access to source and target data during adaptation, whereas SFDA uses only a source-trained model and target data.Source: §4 · p. 6

4.1. Implementation details

• Experiments use Faster R-CNN with ImageNet-pretrained ResNet-50; evaluation reports target-domain mAP at IoU 0.5.Images are resized to a 600-pixel shorter side, and the teacher EMA momentum and pseudo-label confidence threshold are 0.9.Source: §4.1 · p. 6

4.2.1 Adaptation to adverse weather

• The method improves over SFOD by 4.3 mAP and HCL by 3.2 mAP on Cityscapes→FoggyCityscapes, while remaining 2.5–3.5 mAP below two UDA methods that use labeled source data.It also improves 3.5 mAP over the mean-teacher self-training baseline.Source: §4.2.1 · pp. 6–7

4.2.2 Realistic to artistic data adaptation

• The method outperforms ADDA by 4.7 mAP and DANN by 0.3 mAP on PASCAL-VOC→Clipart, and SWDA by 2.6 mAP and I3Net by 1.5 mAP on PASCAL-VOC→Watercolor.It also consistently outperforms pseudo-label self-training and mean-teacher baselines in every category for both adaptations.Source: §4.2.2 · p. 7

4.2.3 Synthetic to real-world adaptation

• On Sim10K→Cityscapes, the method achieves state-of-the-art SFDA performance, improving by approximately 1 mAP over SFOD.The adaptation uses only the car category; the authors report that GCL learns discriminative positive pairs for different cars.Source: §4.2.3 · p. 7

4.2.4 Cross-camera adaptation

• For KITTI→Cityscapes, the authors report that their method outperforms existing state-of-the-art UDA methods.The supplied results text gives no specific margin or comparison with other SFDA methods.Source: §4.2.4 · p. 7

4.3. Ablation analysis

• Strong-student/weak-teacher augmentation consistently improves results over weak-weak and strong-strong variants; graph distillation raises performance from 34.3 to 35.9 mAP, and graph-guided contrastive learning adds 1.9 mAP.The authors attribute the augmentation result to student training on strong views and teacher pseudo-labels from weak views.Source: §4.3 · p. 8, Table 5 · p. 8

5. Conclusion

• The approach uses a contrastive graph loss and an instance relation graph network to model proposal-instance relations and guide positive/negative proposal pairing for contrastive learning.The graph network is built on graph convolution, and the learned relations guide contrastive learning on target representations.Source: §5 · p. 8
• Experiments on multiple detection benchmarks show adaptation performance above state-of-the-art SFDA and many UDA methods.Source: §5 · p. 9

Ablation analysis for different loss functions

• In Cityscapes→FoggyCityscapes, supervised contrastive learning on mean-teacher improves performance by 1.9 mAP; distillation lowers it from 36.2 to 35.9 mAP; graph-guided contrastive learning adds 1.9 mAP on top of GDL.Strong augmentation uses color jitter, grayscale, Gaussian blur, and erasing; weak augmentation uses horizontal flips. Strong and weak augmentation simulate the domain gap between student and teacher.Source: Supplementary material: Instance… · p. 13

Ablations analysis on IRG network

• Learnable IRG modeling of positive relations performs considerably better than K-means pairing, while freezing IRG and using original edge weights yields 36.2 mAP.Source: Supplementary material: Instance… · p. 13, Table 7 · p. 13

Positive proposals visualization

• In the proposal visualization, blue marks the query, green marks positive proposals, and red marks negative proposals.The figure visualizes pairs generated by IRG.Source: Figure 7 · p. 14, Supplementary material: Instance… · p. 14

Detection visualizations

• Detection visualizations report that the model tackles negative transfer and produces high-confidence boxes, while RPN may fail to propose tiny or occluded objects under heavy domain shifts.Instances without proposals cannot benefit from the proposed contrastive learning strategy.Source: Figure 9 · p. 16, Figure 10 · p. 17