When RGB Becomes Unreliable, InfraNet Uses Asymmetric Learning for RGB‑IR Detection

Background and Motivation

When dealing with RGB‑IR object detection, the common assumption is that the two modalities are complementary and always available. The authors observed on the M3FD night set that RGB features become noisy while IR remains clean, and some fusion models even introduce extra activations.

Rethinking the Assumptions

They reformulate the problem: if the auxiliary modality degrades, should the network still treat it as reliable input? InfraNet separates the two assumptions: RGB can provide structural cues during training, but its presence at inference is optional.

InfraNet Architecture

The paper trains two networks instead of a single model that switches modes:

InfraNet‑IR uses paired RGB‑IR images during training, discards the RGB branch and all fusion modules at inference, yielding a pure infrared detector.

InfraNet‑RGB‑IR retains both branches for scenarios where both modalities are reliable.

The core module is QualGate, which predicts a scalar control signal q for each fusion location. q is learned end‑to‑end from the detection loss and drives two synchronized actions: when q is low, RGB contribution is suppressed; simultaneously, IR injection is amplified.

Ablation and Robustness Results

On LLVIP with YOLOv8, the full QualGate achieves 68.8 mAP; removing IR amplification drops to 67.2, removing RGB suppression to 68.4, and fixing q =1 to 67.7. Under strong fog, the quality‑aware model keeps 97.1 AP50 / 67.8 mAP, whereas the naive fusion falls to 92.6 AP50 / 61.8 mAP. The authors note that q is not a generic image‑quality score but a task‑specific control that limits severely degraded RGB from harming training.

Quantitative Evaluation

InfraNet‑IR improves over an IR‑only baseline by 1.4 AP50 and 1.8 mAP without increasing parameters or inference time. Evaluations on four datasets report:

LLVIP: RGB‑IR 70.5 mAP vs. IR‑only 68.8 mAP

FLIR‑Aligned: 89.5 mAP50 (RGB‑IR) vs. 88.1 mAP50 (IR‑only)

DroneVehicle: YOLOv12 InfraNet‑IR reaches 84.7 AP50 / 63.9 mAP

M3FD: consistent gains across lighting conditions

Qualitative Findings

Visual comparisons on M3FD show InfraNet‑IR reduces duplicate boxes and background false alarms compared with SuperFusion, TarDAL, CFT, and FusionMamba. Its high‑response regions focus more tightly on targets rather than roads, lights, or distant structures.

Conclusion

InfraNet does not deny the value of RGB; it simply questions the default symmetric reliance on both modalities. By treating RGB as a controllable auxiliary signal, the method adds a new design axis—reliability versus availability—making infrared detection more robust under adverse conditions.