Research map / Label-efficient and robust learning / Recognition and long-tail
Bidirectional Logits Tree: Pursuing Granularity Reconcilement in Fine-Grained Classification
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 Bidirectional Logits Tree (BiLT), which chains hierarchy-aware classifiers from finest to coarsest granularity plus an adaptive intra-granularity difference learning term, to reconcile granularity competition in fine-grained classification.
Problem
Granularity competition in fine-grained visual classification: when multi-granularity classifiers share one encoder, easier coarse labels dominate feature learning and hurt fine-grained accuracy and mistake severity.
Approach
Instead of parallel independent classifiers on shared features, build classifiers sequentially from finest to coarsest so coarser logits are computed from finer logits (with gradients flowing back), and learn an adaptive intra-granularity distance matrix that calibrates predictions and drives label smoothing.
Key results
- On FGVC-Aircraft, BiLT reaches 81.23% Top-1 accuracy and 2.00 Mistake Severity, versus 79.35% and 2.12 for cross-entropy.
- On CIFAR-100, BiLT achieves the lowest Mistake Severity (2.17) but Top-1 accuracy (77.77) is below HiE (78.63), a failure case acknowledged by the authors.
- Ablation on FGVC-Aircraft shows BiLT gives the largest single-component gain (Top-1 79.35 to 81.07), and all three components together reach 81.34%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace independent shared-feature hierarchy classifiers with BiLT (ablation) | fine-grained classification · FGVC-Aircraft | Top-1 Accuracy: 79.35 -> 81.07 | Helped |
| add AIGDL alone (ablation) | fine-grained classification · FGVC-Aircraft | Mistake Severity: 2.12 -> 2.05 | Helped |
| add LabelSmoothing alone (ablation) | fine-grained classification · FGVC-Aircraft | Top-1 Accuracy: 79.35 -> 75.84 | Hurt |
| add AIGDL to BiLT + LabelSmoothing (ablation) | BiLT + LabelSmoothing · FGVC-Aircraft | Mistake Severity: 1.96 -> 1.95 | Helped |
| add LabelSmoothing to BiLT + AIGDL (ablation) | BiLT + AIGDL · FGVC-Aircraft | Mistake Severity: 1.99 -> 1.95 | Helped |
| add BiLT to AIGDL + LabelSmoothing (ablation) | AIGDL + LabelSmoothing · FGVC-Aircraft | Top-1 Accuracy: 80.47 -> 81.34 | Helped |
| use BiLT + AIGDL + LabelSmoothing rather than strongest listed baseline | fine-grained classification · FGVC-Aircraft | Top-1 Accuracy: 80.72 +/- 0.5849 -> 81.23 +/- 0.5820 | Helped |
| use BiLT + AIGDL + LabelSmoothing rather than strongest listed baseline | fine-grained classification · CIFAR-100 | Top-1 Accuracy: 78.63 +/- 0.3101 -> 77.77 +/- 0.4140 | Hurt |
| use BiLT + AIGDL + LabelSmoothing rather than strongest listed baseline | fine-grained classification · iNaturalist2019 | Top-1 Accuracy: 71.43 +/- 0.2584 -> 70.98 +/- 0.0450 | Hurt |
| use BiLT + AIGDL + LabelSmoothing rather than strongest listed baseline | fine-grained classification · tieredImageNet-H | Mistake Severity: 6.85 +/- 0.0306 -> 6.77 +/- 0.0371 | Helped |
| set α within 0.5 to 1.0 rather than a large value (ablation) | BiLT · FGVC-Aircraft | Mistake Severity | Helped |
| set α to a large value, reducing coarse-level loss weight (ablation) | BiLT · FGVC-Aircraft | Mistake Severity | Hurt |
Surprising findings
- Label smoothing alone reduces Top-1 accuracy on FGVC-Aircraft (75.84 vs 79.35 for cross-entropy), though it lowers hierarchical distance metrics in some settings.
Limitations
- On CIFAR-100 and iNaturalist2019, BiLT does not win on Top-1 accuracy or Hier Dist@1 (e.g., HiE is best on CIFAR-100 Top-1).
- Sensitivity to hyperparameters alpha, beta, epsilon and gamma requires tuning, with performance degrading at extreme values.
Experimental setup
Datasets: FGVC-Aircraft CIFAR-100 iNaturalist2019 tieredImageNet-H
Benchmarks: FGVC-Aircraft CIFAR-100 iNaturalist2019 tieredImageNet-H
Metrics: Mistake Severity Hier Dist@1 Hier Dist@5 Hier Dist@20 Top-1 Accuracy
Baselines: Cross-Entropy HXE Soft-Labels Flamingo CRM HAF HAFrame HiE
Builds on: Conditional Risk Minimization hierarchical cross-entropy soft-label hierarchical training Equiangular Tight Frame hierarchy-aware classification
Closest papers by meaning
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.86)
- Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.85)
- Scaled Decoupled Distillation (CVPR 2024 · similarity 0.84)
- Learning from Neighbors: Category Extrapolation for Long-Tail Learning (CVPR 2025 · similarity 0.84)
- Twofold Debiasing Enhances Fine-Grained Learning with Coarse Labels (AAAI 2025 · similarity 0.83)
- Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification (NeurIPS 2026 · similarity 0.83)
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.