Starting from Zero: A No-Pretraining Object Detectors

Abstract We introduce an innovative object detector that has been trained entirely from scratch , designed to address the shortcomings of contemporary models in terms of scalability and generalization. Unlike state-of-the-art detectors that rely on networks pre-trained on large datasets like ImageNet and COCO, our model steers clear of the potential biases and constraints such training imposes. Building on the foundation laid by DSOD, we present a model that not only improves upon precision but also offers enhanced extensibility, enabling continuous refinement of both accuracy and speed.Our model also introduces a new local attention mechanism that emphasizes salient features while suppressing the less relevant ones, thereby significantly enhancing performance. This plug-and-play module computes tensor weights that can be seamlessly integrated into any network architecture, offering an instant boost in the capability of the model.Starting from ground zero, we trained our model without the crutches of pre-trained weights, an approach that has historically been considered audacious. Yet, our experiments have borne fruit, yielding a model that outperforms DSOD in real-time detection speed while also achieving a mean Average Precision (mAP) of 69.7% on the VOC2007 and 69.8% on the VOC2012 benchmark. This exemplary performance demonstrates the viability and potential of our training paradigm, paving the way for future research into self-sustaining learning frameworks that prioritize adaptability and continuous improvement.