| layout | default |
|---|---|
| title | Semantic Segmentation |
This challenge focuses on zero-shot semantic segmentation for off-road environments, an area with significant potential for innovation due to the complexity and diversity of unstructured terrains. Unlike traditional segmentation tasks, participants will not have pre-labeled training data from the provided dataset and must develop or adapt models to segment scenes without relying on annotated data.
Participants are tasked with segmenting off-road scenes into meaningful categories such as rocks, vegetation, mud, water, and other terrain types. Given the lack of labeled training data, this challenge emphasizes the use of methods such as:
- Transfer Learning: Utilizing models pre-trained on related tasks (e.g., urban driving or indoor segmentation) and fine-tuning them for off-road segmentation.
- Synthetic Data: Generating synthetic training data that simulates off-road environments.
- Unsupervised Learning: Developing segmentation algorithms that can learn meaningful representations of off-road environments without annotations.
- Use of Pre-Trained Models: Leveraging models trained on similar tasks (e.g., outdoor scene understanding) and applying them to the off-road domain.
Participants will receive a dataset containing:
- High-resolution RGB images of various off-road terrains (e.g., forests, deserts, rocky paths).
- LiDAR point clouds for depth and structure understanding.
- Thermal images to capture heat variations across terrains and objects.
Participants will submit their labeled images along with a brief report explaining their approach and any unique strategies they employed.
Submissions will be evaluated on:
- Accuracy: How closely the segmented regions match the hidden ground-truth labels.