YXNiu/Human-Search-Traces

Human Search Traces

Event-level human visual-search annotations for EviViT

Paper · Code · Load the Data · Citation

Human searches record more than where an answer was found: they also record how the annotator explored the scene. This dataset releases 1,144 de-identified human visual-search sessions used to train EviViT, preserving the original event-level records.

At a Glance

Property	Description
Records	1,144 distinct sessions
Source questions	VisualProbe training set
Human annotations	Ordered mouse-interaction events and final evidence boxes
Split / configuration	train / raw
Data file	data/train_search_traces.jsonl
Images	Not redistributed; obtain from the original source
Generated process text	Not included
Annotation license	CC BY-NC 4.0

The questions and image identifiers originate from VisualProbe; the mouse interaction traces and final evidence boxes are our annotations. EviViT uses these interaction records to supervise question-conditioned visual evidence allocation. The original annotation records are preserved, with the workspace-specific prefix removed from each image path.

Loading

pip install datasets

from datasets import load_dataset

sessions = load_dataset("YXNiu/Human-Search-Traces", "raw", split="train")
print(len(sessions))  # 1144

session = sessions[0]
print(session["sample_id"])
print(session["image"])  # filename, not embedded image pixels
print(len(session["events"]))

For full-fidelity processing, you can download and stream the JSONL directly:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="YXNiu/Human-Search-Traces",
    repo_type="dataset",
    filename="data/train_search_traces.jsonl",
)
with open(path, encoding="utf-8") as handle:
    for line in handle:
        session = json.loads(line)
        # Process the ordered session["events"].

Event arrays can be long; the Dataset Viewer preview is not a complete trajectory visualization.

Record Structure

Each JSONL line contains one session. Fields include:

Field	Meaning
id, sample_id	Stable session and source-sample identifiers
image	Source image filename only
question, answer	VisualProbe question and reference answer
events	Ordered interaction records, including pointer/hover, zoom, and box operations when present
final_bboxes_pixels	Final evidence boxes in pixel coordinates
annotation_time	Relative session duration
task_evaluation	Annotation-task evaluation metadata
source_format, source_dataset	Source provenance

Events include a type and relative t, with operation-specific fields such as x, y, or bbox. Event types include hover_move, zoom_start, zoom_move, zoom_end, zoom_reset, create_start, and create_end. Operations can repeat or be absent; the overview is schematic, not a fixed event sequence.

These are mouse-interaction traces, not eye-gaze measurements or human-written chains of thought. Time values are relative to the start of annotation, not wall-clock timestamps. Coordinates are recorded in pixels; interpret spatial changes together with the operation and zoom history.

Getting the Images

Images are not redistributed here. Download them separately from VisualProbe_train, subject to its terms. Use the image filename (for example, visual_probe_train_1001.jpg) to match a session to its source image. Trace analysis can use the JSONL without downloading images.

This is a training-set release, not a train/validation/test split. Do not use VisualProbe Easy, Medium, or Hard evaluation questions as training examples when reproducing the paper. Preparing EviViT component-training inputs also requires the corresponding targets, images, and host features; this raw file is not a precomputed feature bundle.

Provenance, Quality, and Limitations

• The 1,144 records are distinct sessions. One session has no final evidence box; tasks requiring a terminal box should exclude it.
• Some boxes or trace crops require task-specific quality filtering.
• Pointer velocity and acceleration may jump across zoom transitions. Use event types and view history rather than assume a globally continuous motion path.
• The traces reflect the annotation interface and sampled VisualProbe questions; they are not a universal model of human attention.

The HaPRL code repository contains 1,104 overlapping examples in its training files, including model-generated round-by-round textual descriptions used by HaPRL. This dataset releases the full 1,144 event-level training sessions used by EviViT; no generated process text is included here.

Privacy and License

The released schema contains no annotator name, email, account identifier, device identifier, IP address, or wall-clock collection timestamp. These are nonetheless behavioral interaction records: do not attempt to identify or profile annotators. Please report privacy or data-quality issues through EviViT GitHub Issues.

Our human annotations are released under CC BY-NC 4.0. Attribute EviViT and the original VisualProbe/Mini-o3 dataset when using these records. Original VisualProbe content remains subject to its upstream license and is not relicensed by this release. Commercial uses are not permitted under the annotation license.

Citation

Please cite EviViT for the human-search annotations:

@misc{niu2026evivit,
  title         = {EviViT: Evidence-Adaptive Vision Transformers for Fine-Grained Perception},
  author        = {Yaoxin Niu and Zhangquan Chen and Yang Zhang and Xiang An and Zhumei Wang and Chih-Ting Liao and Hongkun Cao and Ruqi Huang},
  year          = {2026},
  eprint        = {2609.37123},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.37123}
}

Also cite Mini-o3 for the VisualProbe questions and images.