Benchmarking Deep Learning-Based Object Detection Models on Feature Deficient Astrophotography Imagery Dataset
Shantanusinh Parmar
TL;DR
The paper tackles object detection in feature-deficient astrophotography by introducing MobilTelesco, a smartphone-captured dataset characterized by extreme noise and sparse target features. It benchmarks seven diverse detectors (including SSD300, Faster R-CNN, PP-YOLOE+, NanoDet+, Sparse RCNN, RetinaNet, and Yolov12) on MobilTelesco, revealing substantial performance degradation compared with COCO-era benchmarks and showing that increased model complexity does not reliably close the gap. Yolov12x performs best among the tested models on MobilTelesco, but still trails COCO performance, highlighting a fundamental mismatch between standard detectors and sparse astronomical imagery. The work motivates future research in data preprocessing (denoising), synthetic data pretraining, segmentation-based or constellation-aware methods, and domain-specific supervision to enable robust, resource-efficient astronomical object detection under extreme conditions.
Abstract
Object detection models are typically trained on datasets like ImageNet, COCO, and PASCAL VOC, which focus on everyday objects. However, these lack signal sparsity found in non-commercial domains. MobilTelesco, a smartphone-based astrophotography dataset, addresses this by providing sparse night-sky images. We benchmark several detection models on it, highlighting challenges under feature-deficient conditions.