DSOD: Learning Deeply Supervised Object Detectors from Scratch (1708.01241v2)

Abstract: We present Deeply Supervised Object Detector (DSOD), a framework that can learn object detectors from scratch. State-of-the-art object objectors rely heavily on the off-the-shelf networks pre-trained on large-scale classification datasets like ImageNet, which incurs learning bias due to the difference on both the loss functions and the category distributions between classification and detection tasks. Model fine-tuning for the detection task could alleviate this bias to some extent but not fundamentally. Besides, transferring pre-trained models from classification to detection between discrepant domains is even more difficult (e.g. RGB to depth images). A better solution to tackle these two critical problems is to train object detectors from scratch, which motivates our proposed DSOD. Previous efforts in this direction mostly failed due to much more complicated loss functions and limited training data in object detection. In DSOD, we contribute a set of design principles for training object detectors from scratch. One of the key findings is that deep supervision, enabled by dense layer-wise connections, plays a critical role in learning a good detector. Combining with several other principles, we develop DSOD following the single-shot detection (SSD) framework. Experiments on PASCAL VOC 2007, 2012 and MS COCO datasets demonstrate that DSOD can achieve better results than the state-of-the-art solutions with much more compact models. For instance, DSOD outperforms SSD on all three benchmarks with real-time detection speed, while requires only 1/2 parameters to SSD and 1/10 parameters to Faster RCNN. Our code and models are available at: https://github.com/szq0214/DSOD .

Overview of DSOD: Learning Deeply Supervised Object Detectors from Scratch

The paper presents a framework known as Deeply Supervised Object Detector (DSOD), which is designed to learn object detectors from scratch, rather than relying on pre-trained networks from large-scale datasets like ImageNet. DSOD addresses several limitations inherent in using pre-trained networks for object detection, such as learning bias due to discrepancies in loss functions and category distributions, as well as challenges in domain transferability, like from RGB to depth images.

Key Contributions

1. Training Object Detectors from Scratch:
   • DSOD is structured to train from scratch, avoiding the use of pre-trained models. This approach circumvents the constraints of transferring pre-trained models across different domains and tasks.
2. Design Principles for Efficient Training:
   • The framework incorporates dense layer-wise connections for deep supervision. This methodology strengthens supervision at earlier network stages, enhancing overall detector performance on limited training data.
3. Compact Model Architecture:
   • DSOD merges dense connections with the single-shot detection (SSD) framework, maintaining real-time speed but with significantly fewer parameters.
   • It achieves state-of-the-art results with about half the parameters of SSD and one-tenth the parameters of Faster R-CNN.

Numerical Results

• DSOD demonstrates superior performance on benchmarks such as PASCAL VOC 2007, 2012, and MS COCO datasets.
• On PASCAL VOC 2007, DSOD achieves 77.7% mAP, outperforming alternative methods such as SSD300 while utilizing fewer parameters.

Theoretical and Practical Implications

• Domain and Task Flexibility:
  • DSOD's ability to train from scratch provides flexibility in adapting to diverse tasks and domains without the limitations imposed by pre-trained models.
• Model Efficiency:
  • The compact nature of DSOD models aids deployment in resource-constrained environments, expanding the applicability of advanced detection techniques to mobile and embedded systems.

Future Directions

• Adapting to Diverse Image Domains:
  • Future work can leverage DSOD to explore object detection in non-traditional image domains, such as multispectral or medical imaging.
• Enhancing Network Efficiency:
  • Developing ultra-efficient DSOD models could further validate its potential for widespread practical applications and set the stage for advancements in Internet-of-Things (IoT) initiatives.

Conclusion

The DSOD framework represents a significant step towards versatile and efficient object detection. By eliminating the dependency on pre-trained models and focusing on effective network design, DSOD addresses key challenges in current detection methodologies, offering robust solutions applicable across diverse domains and tasks. With ongoing improvements and adaptations, DSOD may continue to enhance the capabilities and usability of object detection systems in various applications.