A proposed method for making very small image-training sets achieved the highest reported top-1 accuracy on all three fine-grained benchmarks it tested, at each of the image budgets examined. Top-1 accuracy means the share of cases in which a model's first prediction is correct. With one image per class, DeCO reached 65.53% on CUB-200-2011, 66.04% on FGVC-Aircraft and 63.94% on Stanford Cars.
The results come from a methods study framed around fine-grained dataset distillation, the effort to preserve useful training information in a much smaller dataset. Its design focuses on dense, class-specific evidence spread across different parts of an image, a premise the authors test with compact sets built from selected regions.
Building a small set from local clues
DeCO first scores candidate image patches with a pretrained TransFG teacher. It then spreads the chosen patches across different spatial locations, gathers them into class-specific evidence banks and combines regions from the same class into grid images. The teacher is used while the compact set is being built, not as the model trained on that set.
The experiments covered CUB-200-2011, FGVC-Aircraft and Stanford Cars, using budgets of one, three or five images per class. Images were 224 by 224 pixels. The teacher was a pretrained TransFG ViT-B/16, while the student was a randomly initialized TransFG model. The reported setup used a two-by-two grid and a region area equal to 28% of the image.
The study asks whether localized, spatially diverse, class-specific evidence can be preserved when the number of images per class is limited.
Large gaps in the one-image comparison
At the one-image-per-class budget, DeCO exceeded RDED by 27.28 percentage points on CUB-200-2011, 43.93 points on FGVC-Aircraft and 46.99 points on Stanford Cars under a unified protocol. Compared with the strongest reported baseline under original protocols, the gaps were 10.73, 11.02 and 3.64 points, respectively.
Those comparisons do not all come from the same testing arrangement. RDED was re-evaluated with the shared student setup and region budget. The reported SRe2 L++ and FADRM+ results came from prior work under their original protocols, so the figures combine a matched comparison with comparisons made under different conditions.
A separate matched test compared DeCO with DeCOrand, which kept the same composition and student-training pipeline but chose regions randomly instead of using attention guidance. DeCO consistently outperformed that random-region version across the reported datasets and image-per-class comparisons.
The study also included Grad-CAM visualizations of student attention. Students trained with DeCO generally showed more concentrated responses on discriminative object regions than students trained with RDED.
The evidence has clear boundaries
The reported sensitivity checks found little movement when the teacher's patch size was changed: accuracy shifted by no more than 0.2 percentage points. Varying the depth used to roll out the teacher's attention generally improved accuracy through early and middle depths before becoming stable later. From depth 5 to depth 11, the change was 0.15 points on CUB-200-2011 and 0.08 points on FGVC-Aircraft.
The main results are still point estimates. Results without special markings were averaged over at least three independent runs, but the study reports no error bars, confidence intervals or formal statistical significance tests. The supplied evidence therefore does not establish how the reported differences would fare under formal uncertainty analysis.
An appendix offers an informal analysis of evidence preservation, but explicitly says it is not a formal guarantee of downstream accuracy. The reported evidence is limited to the three named benchmarks, the one, three and five image-per-class budgets, and the specified TransFG teacher and student setup.
The checklist identifies two practical issues that are not explicitly discussed: dependence on a pretrained teacher and the computational cost of constructing the compact dataset. It also reports no dedicated limitations discussion.
What is publicly available
The experiments use publicly available benchmark datasets, but no open-access code repository or public implementation was provided at submission time. Additional implementation and hyperparameter details are stated to be available in the appendix and supplementary material.
The document is an arXiv version 1 preprint dated 26 August 2026. The findings are limited to the three named benchmarks, the one, three and five image-per-class budgets, and the specified TransFG teacher and student setup.
Paper data and sources
Original title: DeCO: Discriminative Evidence Composition for Fine-Grained Dataset Distillation
Authors: Chuixuan Fan, Guang Li, Shijie Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text