Adaptive Temporal Pooling for Object Detection using Dynamic Vision Sensor

Jia Li, Feng Shi, Weiheng Liu, Dongqing Zou, Qiang Wang, Paul-K.J. Park and Hyunsurk Eric Ryu

Abstract

Dynamic Vision Sensor (DVS) is a kind of asynchronous sensor which can capture visual information with low power consumption. Encouraged by the superior properties of DVS, we propose a novel object detection method for DVS. The imaging of DVS is highly related to the object motion speed. For sparse events induced by slow motions, directly applying existing object detection methods cannot obtain reliable features, leading to loss in performance. For this issue, we introduce a new model to extract motion invariant features for object detection using DVS, namely recursive adaptive temporal pooling. The model works in a recursive manner and adaptively integrates features in varied time interval into a holistic representation for current time-step with respect to the motion speed. The pooled features are then fed into a detection sub-network to produce the final results. Our framework is end-to-end trainable and is highly efficient benefitting from the recursive manner and the characteristics of DVS data. Our method is evaluated on a large DVS dataset with about 100k ground-truth images and compared with the state-of-the-art methods.

Session

Posters

Files

PDF iconPaper (PDF)

DOI

10.5244/C.31.40
https://dx.doi.org/10.5244/C.31.40

Citation

Jia Li, Feng Shi, Weiheng Liu, Dongqing Zou, Qiang Wang, Paul-K.J. Park and Hyunsurk Eric Ryu. Adaptive Temporal Pooling for Object Detection using Dynamic Vision Sensor. In T.K. Kim, S. Zafeiriou, G. Brostow and K. Mikolajczyk, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 40.1-40.12. BMVA Press, September 2017.

Bibtex

            @inproceedings{BMVC2017_40,
                title={Adaptive Temporal Pooling for Object Detection using Dynamic Vision Sensor},
                author={Jia Li, Feng Shi, Weiheng Liu, Dongqing Zou, Qiang Wang, Paul-K.J. Park and Hyunsurk Eric Ryu},
                year={2017},
                month={September},
                pages={40.1-40.12},
                articleno={40},
                numpages={12},
                booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
                publisher={BMVA Press},
                editor={Tae-Kyun Kim, Stefanos Zafeiriou, Gabriel Brostow and Krystian Mikolajczyk},
                doi={10.5244/C.31.40},
                isbn={1-901725-60-X},
                url={https://dx.doi.org/10.5244/C.31.40}
            }