Multi-Label Object Detection on the COCO Dataset: A Memory-Efficient CNN Approach for Real-Time Applications
Sarang Pramod Walke, Japjot Singh Kashyap, Vaibhav Raju Sakharwade, Akanksha Raut, Chetan Dhule, Pramod V. Walke · 2025
This paper presents a memory-efficient Convolutional Neural Network (CNN) for multi-label object detection on the COCO dataset, targeting real-time applications. Our model achieves 59.5% [email protected] while processing 72.5 frames per second (FPS) on a single GPU, outperforming Faster R-CNN (45.3 FPS) in speed and YOLOv8 (275MB) in memory efficiency with a 245MB footprint. Key innovations include a memory-mapped data pipeline reducing RAM usage by 40% and lightweight CNN architecture optimized for small-scale objects. Through strategic data augmentation (MixUp, geometric transformations) and class-balancing techniques, the model attains 89.7% accuracy for dominant categories (e.g., people, vehicles), though small-object detection remains challenging (75.3-82%). Comparative analysis demonstrates a 3 x lower memory requirement than Faster R-CNN (16GB) with competitive accuracy, making it viable for edge devices. Future work will integrate Feature Pyramid Networks (FPNs) to address scale variance.