Research map / Detection and segmentation / Open-set and few-shot detection
Supervision Interpolation via LossMix: Generalizing Mixup for Object Detection and Beyond
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Supervision Interpolation and LossMix, which mixes per-image losses instead of labels to regularize object detectors and, combined with Adaptive Teacher, improves cross-domain detection.
Problem
Mixup-style data mixing cannot be applied directly to object detection because of spatial misalignment, foreground/background distinction, and multiple instances; existing union-of-boxes label mixing ignores the mixing ratio.
Approach
Supervision Interpolation generalizes Mixup by interpolating supervision signals; LossMix instantiates it by weighting the detection loss of each mixed image by lambda and (1-lambda), avoiding explicit label/box mixing.
Key results
- On PASCAL VOC with ResNet-50 FPN, LossMix reaches 55.87 AP versus 53.30 baseline and 55.05 Union.
- On MS COCO with ResNet-101 FPN, LossMix reaches 44.07 AP versus 42.28 baseline and 43.87 Union.
- On PASCAL VOC to Clipart1k, LossMix with Adaptive Teacher reaches 51.1 mAP, versus 46.7 for reproduced AT and 49.3 reported in the original AT paper.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace no mixing with LossMix | Faster R-CNN, ResNet-50 + FPN · PASCAL VOC 2007 test | AP: 53.30 -> 55.87 | Helped |
| replace Union with LossMix | Faster R-CNN, ResNet-101 + FPN · PASCAL VOC 2007 test | AP: 55.01 -> 55.91 | Helped |
| replace Union with LossMix | Faster R-CNN, ResNet-50 + FPN · MS COCO val2017 | AP: 41.43 -> 41.82 | Helped |
| replace Union with LossMix | Faster R-CNN, ResNet-101 + FPN · MS COCO val2017 | AP: 43.87 -> 44.07 | Helped |
| add LossMix to Adaptive Teacher | Adaptive Teacher · PASCAL VOC → Clipart1k | mAP: 46.7 -> 51.1 | Helped |
| add LossMix to Adaptive Teacher | Adaptive Teacher · PASCAL VOC → Watercolor2k | mAP: 57.7 -> 59.3 | Helped |
| mix classification and localization losses instead of ROI losses only (ablation) | LossMix on Faster R-CNN, ResNet-50 + FPN · PASCAL VOC | AP: 56.15 -> 56.60 | Helped |
| add classification loss mixing to localization-only LossMix (ablation) | LossMix on Faster R-CNN, ResNet-50 + FPN · PASCAL VOC | AP: 55.31 -> 56.60 | Helped |
| add localization loss mixing to classification-only LossMix (ablation) | LossMix on Faster R-CNN, ResNet-50 + FPN · PASCAL VOC | AP: 56.55 -> 56.60 | No clear effect |
| set Beta mixing parameter α to 0.2 instead of 1.0 (ablation) | LossMix on Faster R-CNN, ResNet-50 + FPN · PASCAL VOC | AP: 56.60 -> 56.21 | Hurt |
| set Beta mixing parameter α to 5.0 instead of 1.0 (ablation) | LossMix on Faster R-CNN, ResNet-50 + FPN · PASCAL VOC | AP: 56.60 -> 56.18 | Hurt |
| add RegMixup-style regular data to LossMix (ablation) | LossMix on Faster R-CNN, ResNet-50 + FPN · PASCAL VOC | AP: 56.60 -> 56.31 | Hurt |
Surprising findings
- Noise mixing, which discards objects from the mixed-in image, still improves over the no-mixing baseline on PASCAL VOC.
- Heavily relying on noise mixing during the adaptation phase of Mean Teacher is reported to bias toward the source domain and underperform AT.
Limitations
- Domain adaptation reproduction of Adaptive Teacher was unstable; the authors had to re-tune hyperparameters and their AT baseline (46.7) is below the reported 49.3.
- Mixing with two images only; mixing three or more was not explored.
- Gains on COCO are mixed; LossMix does not win on AP_S and AP50 in all settings.
Experimental setup
Datasets: PASCAL VOC 2007 PASCAL VOC 2012 MS COCO Clipart1k Watercolor2k
Benchmarks: PASCAL VOC 2007 test MS COCO val2017 Clipart1k Watercolor2k
Metrics: AP AP50 AP75 mAP
Baselines: Baseline (no mixing) Noise mixing Union mixing SCL SWDA DM CRDA HTCN UMT AT
Builds on: Mixup CutMix Adaptive Teacher Faster R-CNN RegMixup
Closest papers by meaning
- Interpolation-Based Semi-Supervised Learning for Object Detection (CVPR 2021 · similarity 0.89)
- Zoom-In-To-Check: Boosting Video Interpolation via Instance-Level Discrimination (CVPR 2019 · similarity 0.87)
- RecursiveMix: Mixed Learning with History (NeurIPS 2022 · similarity 0.86)
- Object-Aware Domain Generalization for Object Detection (AAAI 2024 · similarity 0.86)
- InsCal: Calibrated Multi-Source Fully Test-Time Prompt Tuning for Object Detection (CVPR 2026 · similarity 0.86)
- MUM: Mix Image Tiles and UnMix Feature Tiles for Semi-Supervised Object Detection (CVPR 2022 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.