Object Detection from Scratch with Deep Supervision.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection models often rely on pre-trained networks, leading to learning biases from classification tasks.
• Transferring pre-trained models across domains (e.g., RGB to depth images) presents significant challenges.
• Existing methods for training object detectors from scratch are hindered by limited data and naive backbone structures.

Purpose of the Study:

• To propose Deeply Supervised Object Detectors (DSOD), a novel framework for training object detection models from scratch.
• To address the limitations of pre-trained models in object detection, including learning bias and cross-domain transfer difficulties.
• To establish design principles for effective scratch training of object detectors.

Main Methods:

• Developed DSOD, an object detection framework incorporating deep supervision through layer-wise dense connections.
• Built DSOD upon the single-shot detection (SSD) framework, integrating key design principles for scratch training.
• Evaluated DSOD on PASCAL VOC 2007, 2012, and COCO datasets.

Main Results:

• DSOD achieved superior results compared to state-of-the-art methods on PASCAL VOC and COCO datasets.
• DSOD models are significantly more compact, requiring only half the parameters of the baseline SSD.
• DSOD demonstrated comparable or better performance than Mask RCNN + FPN with substantially fewer parameters and no pre-training.

Conclusions:

• Training object detectors from scratch is feasible and advantageous, overcoming pre-training limitations.
• Deep supervision and layer-wise dense connections are critical for successful scratch training of object detectors.
• DSOD offers a highly efficient and effective alternative for object detection, particularly in resource-constrained scenarios or when domain transfer is required.