Research map / Label-efficient and robust learning / Recognition and long-tail
D3still: Decoupled Differential Distillation for Asymmetric Image Retrieval
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 D3still, a decoupled pairwise similarity differential distillation framework that transfers feature, inconsistent, and consistent differential knowledge from a ResNet101/SwinV2 gallery network to a lightweight query network for asymmetric image retrieval.
Problem
Asymmetric image retrieval, where a lightweight query network must match a large gallery network's embedding space and retrieval order, fails to maintain consistent retrieval ranking under one-to-one pairwise similarity distillation when the query network has limited capacity.
Approach
Replace rigid one-to-one pairwise similarity matching with a distillation loss on relational differences between pairwise similarities in the gallery space, then decouple it into a feature representation loss, an inconsistent-rank differential loss (up-weighted, beta=0.2) and a consistent-rank differential loss (gamma=0.1) so the query network focuses on hard samples.
Key results
- With ResNet18 query and ResNet101 gallery, D3still reaches 64.31% mAP / 79.60% R1 on In-Shop, versus 63.35% mAP for RAML.
- On SOP, D3still achieves 51.33% mAP and 69.18% R1 with the same ResNet18/ResNet101 setup, 1.67% mAP above the best prior distillation method.
- On MSMT17 with ResNet18 query and ResNet101 gallery, D3still gets 18.64% mAP, 1.82% above RAML.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the strongest prior distillation baseline with D3still | ResNet18 query with ResNet101 gallery · In-Shop | mAP (%): 63.35 -> 64.31 | Helped |
| replace the strongest prior distillation baseline with D3still | ResNet18 query with ResNet101 gallery · SOP | mAP (%): 49.66 -> 51.33 | Helped |
| replace the strongest prior distillation baseline with D3still | ResNet18 query with ResNet101 gallery · MSMT17 | mAP (%): 16.87 -> 18.64 | Helped |
| add decoupled differential distillation to feature distillation (ablation) | ResNet18 query with ResNet101 gallery · In-Shop and SOP | mAP (%) and R1 (%): In-Shop mAP: 61.55 -> 63.88; R1: 75.37 -> 78.92; SOP mAP: 49.46 -> 51.00; R1: 66.55 -> 68.65 | Helped |
| add the coupled pairwise differential loss alone rather than feature distillation (ablation) | ResNet18 query with ResNet101 gallery · In-Shop and SOP | mAP (%): In-Shop: 61.55 -> 0.85; SOP: 49.46 -> 25.73 | Hurt |
| add pairwise similarity distillation to feature distillation (ablation) | ResNet18 query with ResNet101 gallery · In-Shop and SOP | mAP (%) and R1 (%): SOP mAP: 49.46 -> 49.96; R1: 66.55 -> 66.94 | Helped |
| add inconsistent differential distillation only (ablation) | ResNet18 query with ResNet101 gallery · In-Shop and SOP | mAP (%): In-Shop: 61.55 -> 63.38; SOP: 49.46 -> 50.57 | Helped |
| add consistent differential distillation only (ablation) | ResNet18 query with ResNet101 gallery · In-Shop and SOP | mAP (%): In-Shop: 61.55 -> 63.46; SOP: 49.46 -> 50.89 | Helped |
| add both consistent and inconsistent differential distillation (ablation) | ResNet18 query with ResNet101 gallery · In-Shop and SOP | mAP (%): In-Shop: 63.46 -> 63.88; SOP: 50.89 -> 51.00 | Helped |
| amplify the inconsistent-sample loss with the decoupled differential loss (ablation) | ResNet18 query with ResNet101 gallery · In-Shop and SOP | mAP (%): In-Shop: 63.88 -> 64.31; SOP: 51.00 -> 51.33 | Helped |
| vary β, the inconsistent differential knowledge weight (ablation) | D3still · In-Shop, SOP, and MSMT17 | mAP (%) | No clear effect |
| vary γ, the consistent differential knowledge weight (ablation) | D3still · In-Shop, SOP, and MSMT17 | mAP (%) | No clear effect |
Surprising findings
- Pure relationship distillation (RKD, PKT) nearly fails in asymmetric retrieval, with mAP near 0.1% on SOP, because feature alignment is missing.
- CSD and ROP perform far worse than pure feature distillation (FitNet), e.g., CSD is 12.05% mAP lower than FitNet on SOP.
Limitations
- Focus limited to ranking properties in image retrieval; applicability to detection or classification may be limited.
- Assumes a significant representational capacity gap between teacher and student networks.
Experimental setup
Datasets: In-Shop Clothes Retrieval Stanford Online Products MSMT17
Benchmarks: In-Shop Clothes Retrieval Stanford Online Products MSMT17
Metrics: mAP R1 FLOPs
Baselines: FitNet CCKD CSD RAML ROP RKD PKT
Builds on: AML ROP CSD RAML FitNet
Closest papers by meaning
- Let All Be Whitened: Multi-Teacher Distillation for Efficient Visual Retrieval (AAAI 2024 · similarity 0.83)
- DREAM: Decoupled Discriminative Learning with Bigraph-aware Alignment for Semi-supervised 2D-3D Cross-modal Retrieval (AAAI 2025 · similarity 0.83)
- Retrieval-Augmented Score Distillation for Text-to-3D Generation (ICML 2024 · similarity 0.83)
- Decoupled Kullback-Leibler Divergence Loss (NeurIPS 2024 · similarity 0.82)
- Asymmetric Mutual Alignment for Unsupervised Zero-Shot Sketch-Based Image Retrieval (AAAI 2024 · similarity 0.82)
- Learning Compatible Multi-Prize Subnetworks for Asymmetric Retrieval (CVPR 2025 · similarity 0.82)
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.