IRIS: A Real-World Benchmark for Inverse Recovery and Identification of Physical Dynamic Systems from Monocular Video

IRIS is a real-world benchmark of 240 high-resolution (3840×2160, 60 fps) videos spanning 8 physical dynamics classes, covering both single-body and multi-body systems. Each setting is recorded with repeated takes and ships with independently measured ground-truth physical parameters, enabling standardized evaluation of inverse parameter recovery and equation-family identification from monocular video.

Introduced in our ECCV 2026 paper. Project page: https://kurbanintelligencelab.github.io/iris-benchmark.github.io/

Dataset at a glance

• 240 videos = 8 classes × 3 settings × 10 repeated takes
• Resolution: 3840×2160 (4K), 60 fps
• Ground truth: physical parameters per (class, setting) in parameters.json, with mean, std, min, max per parameter

Classes

Type	Class	Settings	Videos	Description
Single	dropping_ball	drop_50, drop_100, drop_150	30	Same ball released from rest under gravity (drop height in cm)
Single	falling_ball	small, mid, big	30	Free-falling balls of different sizes
Single	sliding_cone	cone_45, cone_60, cone_80	30	Cone sliding on an inclined surface (incline angle in °)
Single	pendulum	pendulum_20, pendulum_45, pendulum_90	30	Single pendulum (initial angle in °)
Single	rotation	slow, mid, fast	30	Rotating cone of different speed, fixed camera
Multi	hitting_cones	slow, mid, fast	30	Ball colliding with a pyramid of cones (contact / momentum transfer)
Multi	two_moving_pendulums	pendulum_20, pendulum_45, pendulum_90	30	Two pendulums released from the same angle, colliding
Multi	two_moving_pendulum_one_static	pendulum_20, pendulum_45, pendulum_90	30	One moving pendulum strikes a static one
	Total	24 settings	240

Directory structure

IRIS/
├── parameters.json                 # Ground-truth physical parameters
├── IRIS_images/                    # Representative preview frames (5 per class + figures)
├── Dropping_ball/
│   ├── drop_50/   {01..10}.mp4
│   ├── drop_100/  {01..10}.mp4
│   └── drop_150/  {01..10}.mp4
├── Falling_ball/        {big,mid,small}/        {01..10}.mp4
├── Hitting_cones/       {slow,mid,fast}/        {01..10}.mp4
├── Pendulum/            {pendulum_20,45,90}/    {01..10}.mp4
├── Rotation/            {slow,mid,fast}/        {01..10}.mp4
├── Sliding_cone/        {cone_45,60,80}/        {01..10}.mp4
├── Two_Moving_Pendulums/              {pendulum_20,45,90}/  {01..10}.mp4
└── Two_Moving_Pendulum_One_Static/   {pendulum_20,45,90}/  {01..10}.mp4

Folder names are capitalized (e.g. Dropping_ball); the keys in parameters.json are lowercase (e.g. dropping_ball).

Ground truth (parameters.json)

Each (class, setting) entry maps physical quantities to their measured statistics:

"pendulum": {
  "pendulum_45": {
    "angle":          { "mean": 45.0, "std": 0.0, "min": 45.0, "max": 45.0 },
    "rope_length":    { "mean": 0.50, "std": 0.0, "min": 0.50, "max": 0.50 },
    "camera_to_cable":{ "mean": 0.96, "std": 0.0, "min": 0.96, "max": 0.96 }
  }
}

Settings are fixed by design, so std = 0 and min = max = mean. Distances are in meters, angles in degrees.

Class	Parameters
dropping_ball	drop_height, camera_to_horizontal
falling_ball	drop_height, ball_radius, gravity
sliding_cone	angle, hypotenuse, vertical_cathetus, horizontal_cathetus, camera_to_object, camera_height
pendulum	angle, rope_length, camera_to_cable
rotation	camera_to_object
hitting_cones	ball_to_cones, camera_to_cones
two_moving_pendulums	angle, rope_length_1, rope_length_2, camera_to_pendulum
two_moving_pendulum_one_static	angle, rope_length_1, rope_length_2, camera_to_pendulum

Usage

pip install huggingface_hub
huggingface-cli download rasulkhanbayov/IRIS --repo-type dataset --local-dir ./IRIS

import json
from huggingface_hub import snapshot_download

local = snapshot_download(repo_id="rasulkhanbayov/IRIS", repo_type="dataset")
params = json.load(open(f"{local}/parameters.json"))
print(params["pendulum"]["pendulum_45"])

License

Released under CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0). Free for non-commercial research use with attribution.

Citation

@inproceedings{khanbayov2026iris,
  title     = {{IRIS}: A Real-World Benchmark for Inverse Recovery and Identification
               of Physical Dynamic Systems from Monocular Video},
  author    = {Khanbayov, Rasul and Barhdadi, Mohamed Rayan and
               Serpedin, Erchin and Kurban, Hasan},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}