ExtremeTrack Dataset is now available:   Click the link

Competition Outline

The VISTAC (VISual Tracking in Adverse Conditions) challenge returns for its second edition at the 28th International Conference on Pattern Recognition (ICPR) 2026. The primary goal of the competition is to advance the state of the art in object tracking, particularly in environments where traditional tracking algorithms struggle. While significant progress has been made in object tracking within controlled or well-lit settings, much less research has been conducted on tracking performance under challenging weather conditions such as haze and rain. VISTAC-2 seeks to address this gap by providing a platform for the development of robust tracking algorithms that can handle degraded video quality due to environmental factors like haze, rain, and other adverse conditions.

This challenge introduces a new dataset, ExtremeTrack, designed specifically to help improve tracking under these harsh conditions. The dataset contains 188 real-world videos: 96 videos in hazy conditions and  92 in rainy conditions. These videos will serve as the foundation for evaluating new tracking algorithms and assessing their ability to handle the visual degradation commonly associated with adverse weather.

Dataset Overview:

●      Training Data: A total of 128 videos (66 in hazy conditions and 62 in rainy conditions) will be released for training purposes on march 12 , 2026. These videos come with ground truth annotations for the tracking task.

●      Validation Data: 20 videos (10 in hazy conditions and 10 in rainy conditions) will be provided with ground truth annotations to assist participants in fine-tuning their algorithms.

●      Test Data: 40 additional videos (20 in hazy and 20 in rainy conditions) will be released without ground truth on March 20, 2026. These videos will be used to evaluate the tracking performance of participants in real-world scenarios.

The ground truth annotations for the dataset will be in the form of bounding boxes for each object, where each bounding box is defined by its top-left (X1, Y1) and height-width (h, w) coordinates. This provides the exact location of the tracked object in each frame, allowing for precise evaluation of the tracking results.

Competition Tasks:

The VISTAC-2 competition is organized into three distinct tasks, each focusing on different adverse weather conditions to evaluate the robustness of tracking algorithms using the Qualitative Precision (QP) metric:

1. Task 1 – Hazy-Condition Tracking:
   In this task, participants will train their models exclusively on the hazy subset of the ExtremeTrack dataset. The QP metric will be evaluated only on the hazy test videos, measuring the algorithm’s performance under hazy conditions.

2. Task 2 – Rainy-Condition Tracking:

Participants will train their models using only the rainy subset of the ExtremeTrack dataset. The QP metric will then be computed exclusively on the rainy test videos, assessing performance under rainy conditions.

3. Task 3 – Combined-Condition Tracking:
   This task evaluates algorithms trained on the entire combined ExtremeTrack dataset (both hazy and rainy videos). The QP metric will be calculated on the full test set, testing the algorithm’s ability to generalize across multiple adverse weather conditions.

ExtremeTrack Dataset & Submission Guide

1. Annotation Files

The dataset provides two annotation files:

• ExtremeTrack_train.json

• ExtremeTrack_val.json

Note:
Test annotations follow the same format but are not publicly available.

2. Dataset Folder Structure

Frames are organized by weather condition and dataset split.

HAZY/

   Train_Set/

      video1/

      video2/

   Val_set/

      video3/

RAIN/

   Train_set/

      video4/

   Val_set/

      video5/

Each video is stored in its own folder.

Frames may be stored:

• directly inside the video folder, or

• inside an img/ folder.

Example:

video_name/

   0001.jpg

   0002.jpg

or

video_name/

   img/

      0001.jpg

      0002.jpg

3. Annotation JSON Format

Each annotation JSON file is a dictionary:

{

  sequence_id : sequence_object

}

Sequence ID

Format:

video_name_condition

Condition can be:

• Haze

• Rain

Example:

Car01_Haze

Bike03_Rain

4. Sequence Object Fields

Each sequence contains the following information:

• video_dir
  Name of the video folder.

• img_names
  Ordered list of frame paths.

• gt_rect
  Ground truth bounding box for each frame.

• init_rect
  Initial bounding box (same as the first ground truth box).

• attr
  Optional attributes or tags.

5. Bounding Box Format

All bounding boxes use the format:

[x, y, w, h]

Where:

• x = top-left x coordinate

• y = top-left y coordinate

• w = width

• h = height

All values are in pixels.

6. Frame Alignment

Each frame matches its annotation by index:

frame i  → img_names[i]

frame i  → gt_rect[i]

So each frame has one corresponding bounding box.

7. Submission Instructions

The test set ground truth is hidden.

Participants must submit predicted bounding boxes only.

8. Submission File

Submit a single JSON file named:

ExtremeTrack_submission.json

9. Submission Format

The submission JSON should look like this:

{

  video_name : prediction_object

}

video_name must match the test video folder name.

10. Prediction Object

Each video must contain:

• pred_rect
  List of predicted bounding boxes (one per frame).

Optional:

• confidence
  Confidence score for each frame.

Bounding box format:

[x, y, w, h]

11. Important Rules

Do not include ground truth information in the submission.

Do not include:

• gt_rect

• init_rect

• img_names

The evaluation system will automatically match predictions with the test frames.

12. Example Submission

{

  "BlurCar1": {

    "pred_rect": [

      [250, 168, 106, 105],

      [249, 170, 107, 104]

    ]

  },

  "Goat": {

    "pred_rect": [

      [561, 326, 341, 469],

      [560, 325, 342, 470]

    ],

    "confidence": [0.93, 0.92]

  }

}

**Please note that the proposed model was trained solely on the training set of the ExtremeTrack dataset, and fine-tuning was not permitted.

 Competition Objectives

The ICPR 2026 Competition on Visual Tracking in Adverse Conditions (VISTAC-2) aims to advance the state of the art in robust object tracking by addressing the challenges posed by visually degraded environments such as haze and rain. Building on the success of the first edition of VISTAC, this second version focuses on expanding the dataset, improving evaluation standards, and fostering greater collaboration within the research community.

1. Present and release the newly developed ExtremeTrack dataset, specifically designed for single-object tracking in adverse weather conditions. The dataset contains 188 videos (96 hazy and 92 rainy), accompanied by carefully annotated ground truths for training and validation. This dataset provides a unique opportunity for researchers to develop and test tracking algorithms in real-world, visually degraded environments.

2. Encourage the development of robust and adaptive tracking algorithms capable of handling visibility degradation caused by environmental factors such as haze, rain, and motion blur. The competition aims to highlight the limitations of current approaches and inspire new strategies that can generalize better across diverse and challenging conditions.

3. Define a standardized benchmarking framework for evaluating and comparing deep learning–based tracking algorithms using the provided dataset. By introducing consistent evaluation protocols, VISTAC-2 aims to promote transparency, reproducibility, and meaningful comparison of algorithmic performance across different methods.

For queries and suggestions, contact us at: nvisot.ju.etce@gmail.com