← Back to Insights
Research Accepted · ECCV 2026

IRIS: teaching machines to recover the physics hidden in ordinary video

July 27, 2026 · European Conference on Computer Vision (ECCV) 2026

Rasul Khanbayov, Mohamed Rayan Barhdadi, Erchin Serpedin, Hasan Kurban — Hamad Bin Khalifa University & Texas A&M University

Read the full paper (PDF) ↗ Project Page ↗

Point a camera at a swinging pendulum and, in principle, everything you need to know about it is in the footage: its length, its damping, even the equation that governs its motion. Getting a machine to actually extract that physics from pixels is one of the harder open problems in computer vision — and until now, there was no fair way to measure who does it best. IRIS, our new benchmark accepted at ECCV 2026, changes that.

The problem: everyone grades their own homework

Methods for estimating physical parameters from video have so far been evaluated on non-overlapping synthetic datasets — each paper renders its own simulations and reports its own numbers. The only existing real-world dataset covers single-body systems only, and no established protocol exists for the harder question of identifying the governing equations themselves. The result: published methods cannot be meaningfully compared, and it is unclear how well any of them survive contact with reality.

What IRIS provides

IRIS is a high-fidelity, real-world benchmark built for exactly this task. It comprises 240 real-world videos captured at 4K resolution and 60 fps, spanning both single- and multi-body dynamical systems. Every system is recorded under controlled laboratory conditions, with independently measured ground-truth parameters and uncertainty estimates, and paired with its governing equations.

On top of the data, IRIS defines a standardized evaluation protocol covering five axes: parameter accuracy, identifiability, extrapolation, robustness, and governing-equation selection. For the first time, methods can be compared on the same real footage against physically measured truth.

What the baselines reveal

The paper evaluates multiple baselines, including a multi-step physics-loss formulation and four complementary strategies for equation identification — vision-language-model temporal reasoning, describe-then-classify prompting, CNN-based classification, and path-based labelling. These establish reference performance across all IRIS scenarios and, just as importantly, expose systematic failure modes that point to where the field should go next.

Why it matters

Reliable physics-from-video unlocks applications from robotics and digital twins to scientific discovery from observation alone. A shared, real-world yardstick is the prerequisite for progress — the same role ImageNet played for recognition. The dataset, annotations, evaluation toolkit, and all baseline implementations are publicly released.

Applications & economic impact

The economics of physics-from-video are compelling because the sensor is already everywhere: a camera. Reliable inverse recovery turns commodity video into a measurement instrument, with direct applications across industries:

  • Robotics & digital twins — calibrate simulators directly from camera footage of real equipment, instead of instrumenting hardware with dedicated force, torque, and motion sensors.
  • Industrial monitoring — friction and damping coefficients estimated from ordinary video become early-warning signals for predictive maintenance, with no sensor retrofit cost.
  • R&D de-risking — a shared real-world benchmark makes academic and vendor claims directly comparable before organizations commit budget to a method.

Go deeper

Full paper: Download the PDF — complete methodology, evaluation protocol, and baseline results.

Project page: kurbanintelligencelab.github.io/iris-bench

The official proceedings version will be linked when ECCV 2026 publishes.