OD3: Optimization-free Dataset Distillation for Object Detection

What came before

The only prior detection-specific method, DCOD, still relies on uni-level optimization (model inversion of a trained detector, using pixel/feature regularization) to synthesize images, following SRe2L's inversion paradigm; coreset-selection methods for detection (CSOD, training-free instance- segmentation pruning) select whole real images or patches but typically only reach compression ratios above 20%, far coarser than what classification-style distillation achieves.

The problem

Any optimization-based synthesis (bi-level as in classification DD, or DCOD's uni-level inversion) is costly and, in DCOD's case specifically, is tied to fixed gradients and pixel-specific updates that cannot flexibly add or remove context around objects of different sizes; small objects in particular lack sufficient contextual information once inverted, and no detection-condensation method reaches the very low compression ratios (sub-1%) that classification-style decoupled-synthesis methods do.

The idea

Skip optimization entirely: build each synthetic "canvas" by iteratively copy-pasting real object crops (with their real bounding boxes and classes) onto a blank image at random non-overlapping positions (candidate selection), then run a pretrained "observer" detector on the assembled canvas and discard objects it does not confidently recognize (candidate screening) — an add-then-remove loop that a theorem shows strictly increases an "information density x diversity" objective over an add-only baseline.

How it works

For each of IPD (images-per-dataset) synthetic canvases, repeatedly sample bounding-box crops $\langle b_{ir}, c_{ir}\rangle$ from the real dataset, extend each crop's box by a Scale-Aware Dynamic Context Extension (SA-DCE) amount $\ell_{extension}=(1-\frac{a(o_{ir})-\bar a_{min}}{\bar a_{max}-\bar a_{min}})\times\bar r$ (smaller objects get proportionally more added context), and place the crop at a random position if its IoU with existing canvas content is below overlap threshold $\tau{=}0.6$ (up to $M{=}40$ placement attempts). Once a canvas is assembled, a pretrained observer detector $\theta_{obs}$ (Faster R-CNN-101 for COCO) scores every placed object; objects below confidence threshold $\eta{=}0.2$ are removed. This add-then-remove cycle (Eq. 9, $f_{remove}\circ f_{add}$) is proven (Theorem 1, Appendix E) to reach an information-density-plus-diversity objective $G_2\geq G_1$, i.e. at least as high as never removing anything. At post-evaluation time, a target detector is trained on the final canvases using feature-level soft labels: rather than logit-based KD (found ineffective for detection), a PKD-style (Pearson-correlation-based) channel-normalized FPN feature target is distilled from the observer into the target model's backbone+FPN (Eq. 5-6). Stored artifact: assembled canvas images with real bounding boxes/classes; the observer model's FPN features are used only as a training-time soft label, not stored as part of the compressed dataset itself.

Evidence

Table 1 (MS COCO, Faster R-CNN-101 observer, Faster R-CNN-50 target/evaluator, 4 seeds ± SEM): at IPD 1.0%, OD3 reaches 22.40% mAP / 39.50% mAP50 / 22.90% mAP75 vs. DCOD 12.10% / 24.70% / 10.40% (full dataset 39.80% / 60.10% / 43.30%) — a +10.3 mAP / +14.8 mAP50 / +12.5 mAP75 gain over the prior detection- distillation SOTA, and roughly 5-6x the best coreset baseline (Random 8.30% mAP). At the lowest ratio tested, 0.25%: OD3 12.90% mAP vs. DCOD 7.20% (+5.7). Table 2 (Pascal VOC, Faster R-CNN-50 observer=target, mAP50): at IPD 2.0%, OD3 58.70% vs. DCOD 50.70% (+8.0) vs. full dataset 80.35%; the margin over DCOD grows with compression ratio (+0.6 at 0.5%, +4.7 at 1.0%, +8.0 at 2.0%). Table 3 (ablation, label type, COCO 0.25-1.0%): SA-DCE extended boxes (Ex-BBox) consistently beat plain ground-truth boxes (Bbox) and segmentation-mask labels (Mask) across nearly every metric (e.g. 0.5% IPD: 17.20% vs. 16.60% vs. 15.10% mAP), isolating the context-extension mechanism as a positive contributor though a modest one (+0.4-0.6 mAP over Bbox) relative to the overall gain over DCOD.

Limitations

Relies on ground-truth bounding boxes during synthesis (the paper's own stated limitation), restricting applicability to fully unsupervised or label-scarce settings — the method redistributes real annotated crops rather than synthesizing genuinely novel appearance; the two overlap/confidence thresholds ($\tau{=}0.6$, $\eta{=}0.2$) and placement-attempt budget ($M{=}40$) are fixed hyperparameters set once rather than tuned per dataset; the observer model used for screening and soft-label generation must itself be a detector already trained on (or near) the target distribution, so the method is not fully label-free even though it needs no bi-level optimization; the theoretical guarantee (Theorem 1) only shows the add-then-remove objective dominates add-only, not that the chosen objective itself correlates tightly with downstream detector accuracy.