Temporal Feature Networks for CNN based Object Detection

TL;DR

This paper introduces a novel Temporal Feature Network for CNN-based object detection that leverages temporal information to improve detection accuracy, trained from scratch without pre-training, and evaluated successfully on the KITTI dataset.

Contribution

The paper proposes a new CNN architecture that incorporates temporal features for object detection, trained from scratch, and demonstrates its effectiveness on a standard dataset.

Findings

The temporal feature network outperforms the non-temporal baseline.

Training from scratch without ImageNet pre-training is feasible.

Achieves competitive results on the KITTI dataset.

Abstract

For reliable environment perception, the use of temporal information is essential in some situations. Especially for object detection, sometimes a situation can only be understood in the right perspective through temporal information. Since image-based object detectors are currently based almost exclusively on CNN architectures, an extension of their feature extraction with temporal features seems promising. Within this work we investigate different architectural components for a CNN-based temporal information extraction. We present a Temporal Feature Network which is based on the insights gained from our architectural investigations. This network is trained from scratch without any ImageNet information based pre-training as these images are not available with temporal information. The object detector based on this network is evaluated against the non-temporal counterpart as baseline and achieves competitive results in an evaluation on the KITTI object detection dataset.