By accessing this dataset, you agree to the terms of the TASI Benchmark Data Sharing Agreement (https://s3.amazonaws.com/pedestriandataset.situated-intent.net/TASI+Benchmark+Data+Sharing+Agreement_PSI.pdf). This dataset is intended for academic and non-commercial research purposes only.

YAML Metadata Warning:The task_ids "pedestrian-intent-prediction" is not in the official list: acceptability-classification, entity-linking-classification, fact-checking, intent-classification, language-identification, multi-class-classification, multi-label-classification, multi-input-text-classification, natural-language-inference, semantic-similarity-classification, sentiment-classification, topic-classification, semantic-similarity-scoring, sentiment-scoring, sentiment-analysis, hate-speech-detection, text-scoring, named-entity-recognition, part-of-speech, parsing, lemmatization, word-sense-disambiguation, coreference-resolution, extractive-qa, open-domain-qa, closed-domain-qa, news-articles-summarization, news-articles-headline-generation, dialogue-modeling, dialogue-generation, conversational, language-modeling, text-simplification, explanation-generation, abstractive-qa, open-domain-abstractive-qa, closed-domain-qa, open-book-qa, closed-book-qa, text2text-generation, slot-filling, masked-language-modeling, keyword-spotting, speaker-identification, audio-intent-classification, audio-emotion-recognition, audio-language-identification, multi-label-image-classification, multi-class-image-classification, face-detection, vehicle-detection, instance-segmentation, semantic-segmentation, panoptic-segmentation, image-captioning, image-inpainting, image-colorization, super-resolution, grasping, task-planning, tabular-multi-class-classification, tabular-multi-label-classification, tabular-single-column-regression, rdf-to-text, multiple-choice-qa, multiple-choice-coreference-resolution, document-retrieval, utterance-retrieval, entity-linking-retrieval, fact-checking-retrieval, univariate-time-series-forecasting, multivariate-time-series-forecasting, visual-question-answering, document-question-answering, pose-estimation

PSI 2.0 — Pedestrian Scene Intent Dataset

Dataset Summary

PSI 2.0 (Pedestrian Scene Intent) is an egocentric dashcam video dataset for pedestrian crossing intention prediction. All videos are captured from the driver's perspective, introducing a significant domain shift compared to overhead CCTV datasets commonly used in pedestrian behavior research.

This upload is prepared for ECCV 2026 Workshop Track 3: Anomalous Events in Transportation, where PSI 2.0 serves as an out-of-domain test set for evaluating model generalization. The full dataset contains 204 trainval video clips with dense human annotations, with a 40-video held-out test set (newly human-annotated) arriving mid-May 2026.

Key properties:

Egocentric (dashcam) viewpoint — domain shift from standard overhead CCTV datasets
Multiple human annotators per video with intent labels, bounding boxes, and free-text reasoning
LLaVA-format annotations included for direct use with video-language models

Dataset Structure

PSI2.0/
├── trainval/
│   ├── videos/        # 204 mp4 video clips (~15 seconds each)
│   └── annotations/   # per-pedestrian intent labels, bounding boxes, and reasoning
├── test/              # 40 newly human-annotated videos (arriving mid-May 2026)
├── annotations/
│   ├── llava_pedestrian_intent_trainval.json   # LLaVA-format QA pairs (trainval)
│   └── llava_pedestrian_intent_test.json       # (coming with the test set)
└── scripts/
    ├── generate_llava_annotations.py  # convert PSI annotations → LLaVA format
    └── split_clips_to_frames.py       # extract frames from video clips

Directory details

Path	Description
`trainval/videos/`	204 mp4 clips; videos 0001–0146 from the original PSI 2.0 TrainVal split, 0147–0204 from the original Test split
`trainval/annotations/`	JSON files containing per-pedestrian crossing intent labels, bounding boxes, and multi-annotator reasoning descriptions
`test/`	Placeholder — 40 newly human-annotated dashcam videos arriving mid-May 2026
`annotations/`	LLaVA-format instruction-tuning JSONs generated by `generate_llava_annotations.py`
`scripts/`	Utility scripts for frame extraction and LLaVA annotation generation

Annotation Format

Original PSI annotation schema

Each video has two annotation types under trainval/annotations/:

cognitive_annotation_key_frame/<video_id>/pedestrian_intent.json

Annotators label pedestrian crossing intent at key frames; labels forward-fill to subsequent frames until the next key frame.

pedestrians:
  <track_id>:
    observed_frames: [list of frame indices where pedestrian is visible]
    cv_annotations:
      bboxes: [[x1, y1, x2, y2], ...]   # one bbox per observed frame
    cognitive_annotations:
      <annotator_id>:
        intent:      ["cross" | "not_cross" | "not_sure" | ""]  # per observed frame
        key_frame:   [1 | 0]                                     # 1 = label set here
        description: ["free-text reasoning at key frames", ...]

Intent labels:

Label	Meaning
`cross`	Pedestrian intends to cross in front of the vehicle
`not_cross`	Pedestrian does not intend to cross
`not_sure`	Annotator is uncertain

cognitive_annotation_key_frame/<video_id>/driving_decision.json

Per-frame driving speed and direction decisions from human annotators (e.g., maintainSpeed / goStraight).

cv_annotation/<video_id>/cv_annotation.json

Computer vision annotations: per-frame bounding boxes for all tracked objects (pedestrians, cars, etc.) with object type and track IDs.

LLaVA Format

annotations/llava_pedestrian_intent_trainval.json contains instruction-tuning QA pairs generated via a sliding window over each pedestrian track. The label at the last observed frame of each window is used as the ground truth (forward-filled from key frames).

Window defaults: window_size=90 observed frames, window_step=45.

Example entry:

{
  "id": "video_0001_ped_1_w0-89_annotator_0",
  "video": "trainval/videos/video_0001.mp4",
  "window_start": 0,
  "window_end": 89,
  "conversations": [
    {
      "from": "human",
      "value": "<video>\nObserving frames 0 to 89, will pedestrian ped_1 cross the road in front of the ego vehicle?"
    },
    {
      "from": "gpt",
      "value": "The pedestrian has the crossing intention."
    }
  ]
}

GPT response strings:

Intent	Response
`cross`	`"The pedestrian has the crossing intention."`
`not_cross`	`"The pedestrian does not have the crossing intention."`
`not_sure`	`"It is not sure to tell the pedestrian crossing intention."`

Generating LLaVA Annotations

# Default parameters (window_size=90, window_step=45)
python3 scripts/generate_llava_annotations.py

# Custom sliding window
python3 scripts/generate_llava_annotations.py --window_size 60 --window_step 30

Outputs are written to annotations/ and never modify the original source files.

The trainval split is kept unified in a single JSON. To apply a train/val partition, filter by video field: videos 0001–0110 → train, 0112–0146 → val (following the original PSI 2.0 PSI2_split.json groupings).

Data Split

Split	Videos	Source	Status
`trainval`	204	PSI2.0_TrainVal (146) + PSI2.0_Test (58)	Available
`test`	40	Newly human-annotated dashcam videos	Coming mid-May 2026

Data Use Agreement

By accessing this dataset, you agree to the terms of the TASI Benchmark Data Sharing Agreement.

This dataset is intended for academic and non-commercial research purposes only.

Citation

If you use this dataset, please cite the original PSI paper:

@article{chen2021psi,
  title={PSI: A Pedestrian Behavior Dataset for Socially Intelligent Autonomous Car},
  author={Chen, Tina and Tian, Renran and Domeyer, Joshua and Sherony, Rini and Ohn-Bar, Eshed},
  journal={arXiv preprint arXiv:2112.02604},
  year={2021}
}

License

This dataset upload is released under the MIT License. Please also comply with the license terms of the original PSI dataset.

Downloads last month: 22

Total file size:

11.1 GB

Paper for wszsycn/PSI2.0

PSI: A Pedestrian Behavior Dataset for Socially Intelligent Autonomous Car

Paper • 2112.02604 • Published Dec 5, 2021