TASK-AWARE EVALUATION OF JPEG RECOMPRESSION AND RGB BIT-DEPTH REDUCTION FOR OBJECT DETECTION

Abstract

How does image storage affect object detection? Traditional image quality metrics typically focus on human vision, but computer vision models process images differently. To explore this gap, we evaluated the Microsoft COCO val2017 dataset under two distinct degradation methods: frequency-domain JPEG recompression (quality levels 94 to 25) and amplitude-domain uniform RGB bit-depth reduction (8 bits to 1 bit) stored in 24-bit PNG containers. We processed the degraded images using three distinct architectures: two convolutional networks (YOLOv8n, YOLOv8m) and a Vision Transformer (RT-DETR-L). The results demonstrate that JPEG recompression provides a highly efficient tradeoff; quality level 88 halves the dataset storage size (2.02× compression, reducing to 384 MB) with less than a 1.5% relative drop in mAP50 across all models. Conversely, storing uniformly quantized RGB data in lossless PNG containers proved highly inefficient. While 7-bit quantization preserved baseline accuracy, it artificially inflated the physical file size to 1912 MB. Furthermore, at 3 bits per channel, average mAP50 dropped by over 12%, and all networks effectively failed at the 2-bit and 1-bit thresholds. Crucially, scale-aware metrics revealed that small objects are disproportionately vulnerable to compression artifacts across both CNN and Transformer paradigms, suffering up to a 50% relative accuracy loss. Ultimately, we conclude that standard JPEG compression is significantly more robust for object detection than storing uniformly quantized RGB data in unoptimized containers. These empirical findings highlight that conventional human-centric quality metrics are insufficient for predicting downstream neural network performance. Future research should investigate whether these non-linear degradation trends persist across specialized domain-shifted datasets and modern learned compression algorithms.