While several datasets for autonomous navigation have become available in recent years, they have focused on structured driving environments. This usually corresponds to well-delineated infrastructures such as lanes, a small number of well-defined categories for traffic participants, low variation in the object or background appearance, and firm adherence to traffic rules. Under the purview of domain adaptation, the community has also studied various aspects such as weather changes, time of day, or imaging conditions. We propose the novel dataset for segmentation and domain adaptation in unstructured environments where the above assumptions are mostly unsatisfied.
The challenge will feature
Intel is sponsoring prize money of 1000USD for the winners of each of the 5 challenges.
I. Domain Adaptation Challenge:
Manually annotating new data is effort-intensive and not the ideal solution we would like to rely on each time we want to fine-tune a model to a specific location. We intend to tackle this issue in this challenge - especially in adapting models to the challenging IDD domain from multiple source domains. For the source dataset (S), we introduce significant diversity in different dimensions: Mapillary, Cityscapes, Berkeley Deep Drive, and GTA. We sample around 20,000 images from these four datasets depending on several factors (e.g., number of classes and inherent complexity). The source dataset remains uniform for all the sub-tasks. We consider level 2 (16 classes) and level 3 (26 classes) ids of IDD for label spaces in target datasets (T), which provide closed and open-set domain adaptation opportunities (refer to Figure 2 and Section 3.3 of IDD for more details of levels of hierarchy in class labels). The target dataset changes for each sub-task.
II. Segmentation Challenge:
Results from past challenges have highlighted the relative difficulty of the IDD dataset. The dataset used for previous challenges consisted of 20,000 images, finely annotated with 34 classes collected over 200 drive sequences on Indian roads. The label set was expanded compared to popular benchmarks such as Cityscapes, accounting for the new classes. This task involves semantic segmentation, which are now widespread in the computer vision community.
For more details, refer instructions and benchmark