Detectors for very small objects, such as vehicles a few pixels wide in aerial photos, pests on light-trap boards, or dim infrared targets, almost always start from backbones pre-trained on web-scraped photographs. Two costs come with that. The first is data debt: licensing, privacy, and bias problems. The second is a mismatch in object scale. Web photos mostly show large, centered objects, so nothing in pre-training teaches the network to keep a 6-pixel signal alive as its feature maps are repeatedly shrunk.
Formula-driven supervised learning (FDSL) trains networks only on images drawn from mathematical equations, such as fractals or sums of sine waves, and never uses a real photo. It already matches real-image pre-training on classification. It still trails on detection, and it has never been tested on tiny objects. We propose to render formula shapes deliberately at tiny pixel sizes, scattered among formula-generated clutter, and to pre-train the backbone to find and name them. The claim is that the size distribution and the localization task, rather than the choice of formula or the number of images, are what carry over to tiny-object detection.
Core Claim
Existing formula datasets draw their shapes large, clean, and at a single resolution. Visual Atoms renders sine-wave contours and reaches 83.7% ImageNet top-1 with ViT-Base, against 84.2% for JFT-300M. The same paper notes that a gap remains on detection. SegRCDB generates dense radial contours for segmentation pre-training and found that instance density drives transfer quality. Industrial Synthetic Segment Pre-training (InsCore) uses nested hollow masks and beats a fine-tuned SAM by 6.2 mIoU, but its gains plateau beyond 100k images. None of these papers controls the pixel footprint of the rendered objects. None evaluates on a tiny-object benchmark.
We hypothesize the following. If a backbone is pre-trained to localize formula instances whose footprint follows the tiny-object regime (roughly 2 to 16 pixels on a side), rendered with area-averaging so that they look like sensor-integrated blobs rather than crisp vectors, it will keep tiny signals through its strided stages better than a backbone pre-trained on large formula shapes. That advantage should show up as higher very-tiny-object precision after fine-tuning. We also hypothesize that the InsCore plateau reflects saturated task difficulty rather than exhausted formulas, and that shifting the size distribution toward tinier objects restores scaling with image count.
This proposal differs from InsCore and SegRCDB in treating object scale, and the loss of resolution a network imposes at small scale, as the variable being designed. It differs from MoireDB in using formula clutter as hard negatives inside a pre-training localization task, rather than as a PixMix texture augmentation for classification robustness.
Research Program
The generator, which we will call TinyAtoms, composes three formula sources. The first is foreground instances: VisualAtom contours and radial contours, each class defined by formula parameters. The second is backgrounds: $1/f^\alpha$ noise fields, meaning random textures whose power falls off with frequency the way natural images do, combined with moiré interference built from superposed concentric circles, following MoireDB's formulas. The third is distractors: shapes from a held-out formula family with no label, playing the role of the "massive negatives" that SIRST-5K found essential for suppressing false alarms. SIRST-5K sampled real sensor noise to build its negatives; ours come from formulas alone.
Each instance is rendered at 8× its target size and then box-filtered down, so that its final footprint $k$ pixels (side length) is drawn from a log-uniform distribution on $[k_{min}, k_{max}]$. The default range is $[2, 32]$. This keeps sub-pixel edges and contrast loss physically plausible instead of producing aliased vector shapes. Contrast against the background is also sampled, down to near-invisible levels. Pre-training uses a dense heatmap-plus-class objective at the stride-8 feature level: predict instance centers and formula class. This forces the network to localize small things rather than classify whole images.
The analysis contribution is a diagnostic we call signal survival. We plant a single tiny formula instance in a clutter image. At each backbone stage $\ell$, we fit a linear probe that tries to recover the instance's presence and position from the stage-$\ell$ feature map. Survival $S_\ell(k)$ is the probe's detection AUC as a function of object size $k$. The curve shows where a given pre-training regime lets tiny signals die. We will test whether it predicts downstream very-tiny-object gains across all pre-training variants. A diagnostic that does predict them would also let practitioners screen pre-training choices cheaply.
We use ResNet-50, because the pooling and detection numbers in the file are reported on it, and ViT-S with a detection neck. Pre-training runs at 100k, 1M, and 3M images. The 100k scale makes a direct comparison with InsCore's plateau point.
Baselines
The naive reference is a backbone trained from scratch on each downstream dataset with the same fine-tuning schedule. It tells us how much any pre-training helps at all on tiny objects.
The ethical-sourcing baseline is the same backbone pre-trained on VisualAtom-21k-style data or FractalDB at an equal image count and equal pre-training GPU-hours. It represents the best existing formula pre-training. The file gives no reported result on Lacuna for FractalDB or VisualAtom pre-training on AI-TOD-V2, NUDT-SIRST, or LLPD-26, so we will run these ourselves. To beat it, TinyAtoms should improve AI-TOD-V2 very-tiny AP by a margin larger than the seed-to-seed spread (three seeds). The gain should be concentrated in the smallest size bins.
The real-data ceiling is ImageNet-1k supervised pre-training. It is what practitioners actually use, and it answers whether the ethical route costs accuracy. Matching it on very-tiny AP, while trailing on large-object AP, would already be a useful result. Exceeding it on very-tiny AP would be the strong result.
The architectural competitor is ImageNet pre-training with the pooling replaced by LIP. LIP learns a per-location importance map before downsampling and raised Faster R-CNN small-object AP on COCO from 21.9% to 24.0%. We will also include SoftPool, a parameter-free softmax-weighted pooling operator. These represent fixing the network instead of the data. We run them with the same detector, schedule, and tuning budget, and we also test TinyAtoms combined with LIP.
As field reference points, we report our numbers next to the published ones. FSDN-DETR reports very-tiny AP of 16.20 on AI-TOD-V2 with 50% of the training data, using a COCO-pre-trained DETR plus fuzzy adapters. We will reproduce its 50%-data protocol for a like-for-like comparison. AEDNet reports 82.77% IoU on NUDT-SIRST. MSR-RCNN reports 67.4% AP50 on LLPD-26.
The key ablation, called Big-Atoms, uses the identical TinyAtoms generator, formulas, clutter, objective, and image count, with the size distribution moved to $[32, 256]$ pixels. Any gain that survives this comparison can be attributed to scale rather than to formulas or the localization task. A second ablation removes the area-averaged rendering and uses aliased point sampling instead. A third removes the distractors.
Why not just…?
Why not just downsample existing VisualAtom or FractalDB images during pre-training? This is the most serious threat. We do not dismiss it; we include it as a baseline. The "shrunken VisualAtom" variant resizes whole classification images so that the shape covers 2 to 32 pixels, keeping the classification loss. If it matches TinyAtoms, then the size distribution alone matters and the localization task plus clutter is unnecessary, and we will report that simpler finding. We expect it to fall short for two reasons. It never presents a tiny shape among competing structure, which is the condition SIRST-5K's negative-synthesis gains point to. And an image-level label gives no incentive to keep the shape's position through pooling.
Why not just fix the architecture with LIP, SoftPool, or max-pooling? The pooling literature disagrees with itself. S4ND favors max-pooling for tiny lung nodules, SoftPool argues max-pooling discards information, and LIP argues any fixed operator is suboptimal. That disagreement suggests the answer depends on what the features were trained to represent. Architecture fixes also add compute and still start from web data. Our factorial comparison (pre-training source crossed with pooling operator) shows whether the two are complementary, whether one replaces the other, or whether the pooling winner changes with pre-training data. The last outcome would itself explain the disagreement.
Why not just pre-train on COCO or ImageNet and use copy-paste or ContextMix augmentation at fine-tuning? ContextMix shows that context-preserving resizing beats CutMix on small defects (Macro F1 0.948 versus 0.901). It is therefore a strong fine-tuning augmentation, and we apply it to every arm equally rather than treat it as a rival. The web-data baseline stays in as the real-data ceiling. The point is not that real data is worse but whether it can be dropped from pre-training at little or no cost on tiny objects. DiverSeg found that structural diversity within an image beats raw resolution. That finding is consistent with our bet, but DiverSeg still curates real photos with SAM and Detic.
Proposed Experiments
The main experiment fine-tunes every pre-trained backbone with Faster R-CNN on AI-TOD-V2, using full data and the 50% protocol, and on LLPD-26. It also fine-tunes a segmentation head on NUDT-SIRST and MDvsFA, reporting IoU, nIoU, detection probability, and false-alarm rate. We add COCO small-object AP to check that general detection does not regress. We report precision broken down by object-size bin. A win must come from the tiniest bins to support the claim.
The decisive checks are five. First, TinyAtoms must beat Big-Atoms and shrunken VisualAtom on very-tiny AP at equal image count. Second, the signal-survival curves must separate the variants and rank-correlate with downstream very-tiny AP across all pre-training arms. Third, false-alarm rate on NUDT-SIRST must fall when formula distractors are included; if it does not, formula clutter cannot stand in for real negatives. Fourth, performance at 100k, 1M, and 3M images must keep rising for TinyAtoms where InsCore reported a plateau. Fifth, the pre-training-by-pooling factorial must show whether gains add up.
A negative result on the first check would mean scale-matched formula rendering adds nothing beyond simple resizing. That is still worth publishing, as it would settle a question the ethical-sourcing literature has left open.
References
- Visual Atoms: Pre-Training Vision Transformers with Sinusoidal Waves
- SegRCDB: Semantic Segmentation via Formula-Driven Supervised Learning
- Industrial Synthetic Segment Pre-training
- MoireDB: A Formula-driven Image Dataset for Robustness Enhancement
- SIRST-5K: Exploring Massive Negatives Synthesis With Self-Supervised Learning for Robust Infrared Small Target Detection
- LIP: Local Importance-Based Pooling
- Refining activation downsampling with SoftPool
- FSDN-DETR: Enhancing Fuzzy Systems Adapter with DeNoising Anchor Boxes for Transfer Learning in Small Object Detection
- AEDNet: Anisotropic Enhanced Dual-Branch Network for Infrared Small Target Detection
- MSR-RCNN: A Multi-Class Crop Pest Detection Network Based on a Multi-Scale Super-Resolution Feature Enhancement Module
- S4ND: Single-Shot Single-Scale Lung Nodule Detection
- ContextMix: A context-aware data augmentation method for industrial visual inspection systems
- Rethinking Image Super-Resolution from Training Data Perspectives