vbdai/TAD

Dataset for "From Segments to Scenes: Temporal Understanding in Autonomous Driving via Vision-Language Models"

---

Sample Usage

You can download the benchmark using the following Python code snippet.

from datasets import load_dataset 

# --- Load a specific subset by name --- 
action_duration = load_dataset("vbdai/TAD", name="action_duration") 
exact_answer_action = load_dataset("vbdai/TAD", name="exact_answer_action") 
mc_action = load_dataset("vbdai/TAD", name="mc_action") 
relative_action_localize = load_dataset("vbdai/TAD", name="relative_action_localize") 
temp_action_localize = load_dataset("vbdai/TAD", name="temp_action_localize") 
temp_object_localize = load_dataset("vbdai/TAD", name="temp_object_localize") 
temp_ordering = load_dataset("vbdai/TAD", name="temp_ordering") 
keysegments = load_dataset("vbdai/TAD", name="keysegments") 
raw_labels = load_dataset("vbdai/TAD", name="raw_labels") 
print("
Successfully loaded TAD benchmark from HF Hub")

Then insert the path of the .cache folder including the benchmark splits (e.g. C:\user\.cache\huggingface\datasets\vbdai___tad\) into $BENCHMARK_PATH in the shell files discussed in the following sections (Shell Scripts).

Dataset Structure

TAD-Bench is an autonomous driving QA benchmark that consists of seven tasks (or question types). There is a folder in the dataset that corresponds to each of these seven tasks as follows:

• Action Duration: action_duration
• Exact Answer Action Recognition: exact_answer_action
• Multiple Choice Action Recognition: mc_action
• Relative Temporal Action Localization: relative_action_localize
• Temporal Action Localizaiton: temp_action_localize
• Temporal Object Localization: temp_object_localize
• Temporal Ordering: temp_ordering

Within each of these folders, there is a single parquet containing the list of questions and answers for that question type.

The following table shows the breakdown of TAD-Bench by task type. There are a total of 5,861 questions that are divided relatively evenly between ego vehicle and non-ego vehicle-related questions. All question types are well-represented, with a minimum of 80 questions per category. For complete definitions and descriptions of the tasks, please review the full paper.

In addition to the folders corresponding to the seven question types, two additional folders are provided:

• keysegments
• raw_labels

A unique aspect of TAD-Bench is that it poses questions about short, cropped video segments, as well as entire full-length videos to assess varying aspects of temporal understanding of VLMs. As such, we have defined numbered segments for each of the videos in the dataset along with the frames that belong to each of those segments. This segment and frame information is provided in the parquet file within the keysegments folder.

During the development of TAD-Bench, vehicle action labels were annotated manually for most of the vehicles. These action labels were required in order to form the QAs, since TAD-Bench is a temporal understanding benchmark focused on action-centric tasks. To our knowledge, no prior work has released segment-level action labels for the vehicles in NuScenes. By themselves, these labels may be useful for the community to assess vehicle action recognition. The segment-level action labels for the vehicles in the TAD-Bench videos are provided in a single parquet in the raw_labels folder.

Please note that the TAD-Bench benchmark is built on top of the NuScenes validation split, which contains 150 videos, each roughly 20 seconds in length. The segments that are used in this work are typically around 5 seconds. The NuScenes dataset can be downloaded from the official NuScenes website: https://www.nuscenes.org/

TAD-Bench Additional Statistics

Vehicle Statistics

A breakdown of the vehicle types considered in TAD-Bench along with their counts, is shown in the table below. The benchmark considers all vehicle categories relatively evenly, the lone exception being “Emergency Vehicle”, which is understandably quite rare in the original data itself. More information regarding the vehicle statistics is provided in the full paper.

Action Statistics

Statistics of the vehicle actions used in TAD-Bench are provided in the bar plot below. The data is presented in terms of ego, non-ego, as well as the combined total. For TAD-Bench, the following vehicle actions were considered:

• Change lane to left: Lane (L)
• Change lane to right: Lane (R)
• Starting
• Stopped
• Stopping
• Straight, constant speed: Straight
• Turn left: Turn (L)
• Turn right: Turn (R)

  

The two most dominant action categories are “stopped” and “straight, constant speed”. Notably, the ego vehicle is most commonly found to be traveling “straight, constant speed”; whereas, non-ego vehicles are most often stopped. A total of 4,481 vehicle actions were labeled during the human annotation procedure, which are now provided in the raw_labels folder.

Source Code

Source code and scripts that can be used to evaluate models on TAD-Bench are available at https://github.com/vbdi/tad_bench.

Citation

If this is useful, please cite it in your documents.

@misc{tad_bench,
      title={From Segments to Scenes: Temporal Understanding in Autonomous Driving via Vision-Language Models}, 
      author={Kevin Cannons, Saeed Ranjbar Alvar, Mohammad Asiful Hossain, Ahmad Rezaei, Mohsen Gholami, Alireza Heidarikhazaei, Zhou Weimin, Yong Zhang, Mohammad Akbari},
      year={2025},
      eprint={2512.05277},
      archivePrefix={arXiv},
      primaryClass={CV},
      url={https://arxiv.org/abs/2512.05277}, 
}

References

The benchmark is created based on NuScenes dataset. The tools in nuscenes-devkit are used for loading and analyzing NuScenes dataset. We thank the contributors for their great work!

License

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.