AnoVox [Sample]

The presented AnoVox dataset is a dataset of abnormal scenarios with unknown objects. The benchmark relies on the CARLA 0.9.14 simulator. This dataset includes a collection of road scenarios that feature abnormal objects. These scenarios were developed using the CARLA Simulator. Next to unknown objects, a future release will also include unexpected actor behavior.

The primary repository for this dataset is the "Anovox" directory. It comprises ten distinct scenario folders, each assigned a unique identifier. These folders encapsulate a comprehensive array of data components crucial for characterizing and comprehending abnormal traffic situations. Each scenario vividly presents a specific traffic context and covers a time span of 18.5 seconds, equivalent to 185 ticks.

Within each individual scenario folder, the following subfolders and files are present:

1. "ACTION": This folder contains the action state values attributed to the ego vehicle for every tick throughout the scenario.

2. "ANOMALY": This repository houses anomalous objects or occurrences that take place within the scenario. Additionally, this folder has the potential to incorporate irregular driver and pedestrian actions in future iterations.

3. "DEPTH_IMG": This section comprises depth images that encode the spatial depth of each pixel through the utilization of RGB channels (for more information, refer to the CARLA docs: https://carla.readthedocs.io/en/latest/ref_sensors/#depth-camera).

4. "PCD": This section encompasses point cloud data derived from lidar scans, effectively representing the semantic segmentation of the simulated environment.

5. "RGB_IMG": This section hosts RGB images corresponding to each frame of the scenario.

6. "SEMANTIC_IMG": Here, you will find images that offer ground truth information via semantic segmentation.

7. "SEMANTIC_PCD": This section contains point cloud representations with embedded semantic segmentation details, further enriching the ground truth information.

8. "VOXEL_GRID": This section provides ground truth via a voxel representation of the surroundings.

Additionally, the dataset is augmented by a separate directory titled "Scenario_Configuration_Files." This directory comprises eight JSON files tailored for different urban environments, distinguished by their names ("Town01," "Town02," "Town03," "Town04," "Town05," "Town06," "Town07," "Town10HD"). These JSON files play a crucial role in ensuring the reproducibility of scenarios. They encompass critical details, such as the spawn points of anomalies and the ego vehicle, alongside weather presets.

The color palette used to represent the semantic segmentation is based on the Cityscapes color palette. A detailed description of this color palette can be found in the document "color_palette.txt".