Noise-Robust Tiny Object Localization with Flows

TL;DR

This work tackles the persistent difficulty of tiny object localization under annotation noise. It introduces Tiny Object Localization Flow (TOLF), which models residual localization errors with normalizing flows and employs uncertainty-guided gradient modulation to down-weight noisy supervision. Across AI-TOD, DOTA-v2.0, TinyPerson, and COCO, TOLF yields consistent improvements over strong baselines, notably improving tiny-object performance and demonstrating that non-Gaussian, flow-based error modeling better captures annotation noise. The approach offers a principled, data-driven path to robust tiny object detectors with practical impact on safety-critical vision tasks.

Abstract

Despite significant advances in generic object detection, a persistent performance gap remains for tiny objects compared to normal-scale objects. We demonstrate that tiny objects are highly sensitive to annotation noise, where optimizing strict localization objectives risks noise overfitting. To address this, we propose Tiny Object Localization with Flows (TOLF), a noise-robust localization framework leveraging normalizing flows for flexible error modeling and uncertainty-guided optimization. Our method captures complex, non-Gaussian prediction distributions through flow-based error modeling, enabling robust learning under noisy supervision. An uncertainty-aware gradient modulation mechanism further suppresses learning from high-uncertainty, noise-prone samples, mitigating overfitting while stabilizing training. Extensive experiments across three datasets validate our approach's effectiveness. Especially, TOLF boosts the DINO baseline by 1.2% AP on the AI-TOD dataset.

Figures (5)

• Figure 1: Pathological predictions due to overfitting label noise. (a) Inaccurate ground-truth annotation covering background shadows. (b) Overfitting leads to false positives in background regions that resemble noisy annotations. (c) TOLF exhibits low confidence in uncertain locations and more accurate localization. (d) A noise sensitivity analysis that injects Gaussian noise into training annotations and measures detection performance across object scales. Results reveal tiny objects exhibit the largest degradation. Detection performance is evaluated by training a 1$\times$ FCOS detector tian2019fcos. The model is trained on the AI-TOD wang2021tinytrainval and validated on the AI-TOD test.
• Figure 2: Annotation quality statistics based on manual inspection of 532 bounding boxes from 10 AI-TOD test images. Red boxes/bars represent background regions mistakenly labeled as foreground, yellow boxes/bars indicate inaccurate or loose bounding boxes. The results show that nearly 34.2% of annotations are noisy.
• Figure 3: Overview of the noise-robust localization framework TOLF. The localization head predicts the mean $\hat{T}$ and uncertainty $\hat{\sigma}$ for each bounding box. The normalized prediction error is then modeled by a normalizing flow $G_{\phi}$ to capture complex, non-Gaussian error distributions. This enables robust estimation of the prediction distribution $P_\phi(t)$, which accounts for uncertainty and label noise.
• Figure 4: (a) Flexible distribution modeling enabled by TOLF, which better captures real-world noise compared to Gaussian assumptions. (b) Illustration of noise-robust localization. The model predicts a distribution $P_\theta(t)$ centered around the expected true location instead of regressing to noised Dirac ground-truths. This distributional supervision reduces overfitting to label noise and enabling uncertainty-aware localization.
• Figure 5: Left. Learned 1D error distributions for bounding box coordinates (left, right, top, bottom) using normalizing flows. Each distribution $P_{\theta}(\overline{t})$ plots the residual error for a specific coordinate (e.g., left boundary), conditioned on fixed values of the other coordinates. Right. Average predicted variance $\sigma$ of four coordinates overlaid with input.