The KITTI Vision Benchmark Suite

This paper introduces a lightweight semantic
segmentation model, called SDBNetV2, an improved
version of Short-term Dense Bottleneck Network
(SDBNet) without increasing computational costs.
Relying on Short-term Dense Bottleneck (SDB), the
backbone of the proposed model is designed which
provides varied field-of-views to capture
different geometrical objects in a complex scene.
It also proposes a dense feature refinement and an
improved semantic aggregation modules at the
decoder end to enhance the contextualization and
object localization of the scene. Compared to
SDBNet, the proposed model produces 5% more test
accuracy on KITTI test set while having only 0.1
million more parameters.

1.5 million

@article{singha2023improved,
title={Improved Short-term Dense Bottleneck
network for efficient scene analysis},
author={Singha, Tanmay and Pham, Duc-Son and
Krishna, Aneesh},
journal={Computer Vision and Image
Understanding},
pages={103795},
year={2023},
publisher={Elsevier}
}