Moses Guttmann, CTO and founder of Allegro, presents the "Optimizing SSD Object Detection for Low-power Devices" tutorial at the May 2019 Embedded Vision Summit.
Deep learning-based computer vision models have gained traction in applications requiring object detection, thanks to their accuracy and flexibility. For deployment on low-power hardware, single-shot detection (SSD) models are attractive due to their speed when operating on inputs with small spatial dimensions.
The key challenge in creating efficient embedded implementations of SSD is not in the feature extraction module, but rather is due to the non-linear bottleneck in the detection stage, which does not lend itself to parallelization. This hinders the ability to lower the processing time per frame, even with custom hardware.
Guttmann describes in detail a data-centric optimization approach to SSD. The approach drastically lowers the number of priors (“anchors”) needed for the detection, and thus linearly decreases time spent on this costly part of the computation. Thus, specialized processors and custom hardware may be better utilized, yielding higher performance and lower latency regardless of the specific hardware used.