DGL-MOTS Dataset for Autonomous Driving

1. Introduction

The multi-object tracking and segmentation (MOTS) is a critical task for autonomous driving applications. In this work, we offer the DGL-MOTS Dataset (Figure 1.), which includes 106,089 instance masks for 1,632 distinct objects in 40 video frames. Our effort exceeds the state-of-the-art KITTI MOTS [1] and BDD100K MOTS [2] in terms of annotation quality, data diversity, and temporal representation. Results on extensive cross-dataset evaluations indicate the significant performance improvements of several state-of-the-art methods trained on our DGL-MOTS dataset.

Figure 1. A showcase of the DGL-MOTS dataset. We collect data based on different driving scenarios and organize training data based on different settings in terms of highway, local, parking, and residential.

2. Data Download
We are currently running the annotation corrections of our dataset. Some sample data are available HERE.
Annotation statistics is displayed in Table 1.
Table 1. Annotation statistics. Our dataset outperforms the KITTI MOTS in annotation volume and density. BDD100K offers the largest training data but selected sequentially from video frames, which include redundant temporal information.

3. Raw Data Download
Highway
Local
Parking
Residential

4. DGL-Labeler Github
     Coming Soon ...

5. Reference

[1] Voigtlaender, P., Krause, M., Osep, A., Luiten, J., Sekar, B. B. G., Geiger, A., & Leibe, B. (2019). Mots: Multi-object tracking and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 7942-7951).

[2] Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., ... & Darrell, T. (2020). Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 2636-2645).