Improving Object Detection on Fisheye Images via Panoramic Reprojection Using a Non-Parametric Projection Model (NPPM) and YOLO
TL;DR
The result shows that even with very limited training data, doing the detection in the NPPM-calibrated panoramic domain is enough to deliver a clear and useful improvement, which is particularly relevant for surveillance, traffic monitoring, and aerial patrol applications.
Abstract
Fisheye cameras give a single shot of a very wide scene, but the strong distortion they produce makes things harder for object detectors. CNN-based detectors like YOLO are built for normal perspective images, so when they are trained directly on fisheye frames the accuracy drops, especially for objects near the edge of the lens where the distortion is worst. A common workaround is to unwrap the fisheye image to a panorama before detection, but most existing pipelines assume a fixed equidistant projection that may not match the actual lens. We use a Non-Parametric Projection Model (NPPM), implemented through the MOIL SDK, to perform the unwrapping with the lens’s real projection behavior. The fisheye images and their bounding-box annotations are mapped to an equirectangular panorama, a YOLO26 detector is fine-tuned on the panoramic data, and the predictions are mapped back to the original fisheye view at inference. On a small custom dataset of UAV-captured fisheye images with three classes (car, motorcycle, person), the proposed pipeline raises mAP@0.5 to 0.793 ± 0.005 and mAP@0.5:0.95 to 0.556 ± 0.013, compared to 0.667 ± 0.015 and 0.439 ± 0.006 for a YOLO trained directly on the raw fisheye and to 0.680 ± 0.005 and 0.485 ± 0.009 for an equidistant-panorama baseline that uses no NPPM-calibrated information (three random seeds per configuration; Welch’s t-test p < 0.01 on both metrics for every pairwise comparison involving NPPM). The equidistant baseline by itself does not move recall over raw, while NPPM gains +0.143 recall (Cohen’s d = 7.8) over it, indicating that the panoramic layout alone is not sufficient, the lens-fitted projection is what drives the improvement. The result shows that even with very limited training data, doing the detection in the NPPM-calibrated panoramic domain is enough to deliver a clear and useful improvement, which is particularly relevant for surveillance, traffic monitoring, and aerial patrol applications.