By the Vision & Graphics Laboratory, USC Institute for Creative Technologies

Featuring work by Bipin Kishore, Research Electrical Engineer, USC ICT (Originally published November 2nd, 2023)

A light stage uses structured lighting and a multi-camera setup to capture shape, texture, reflectance, and motion. The Light Stage facial capture pipeline depends on tight photometric accuracy across dozens of synchronized cameras. For years, USC ICT’s Vision & Graphics Lab ran this pipeline on hi-speed DSLRs (the Canon 1DX) and Ximea machine vision cameras, using them for both multi-view stereo (MVS) and One Light at a Time (OLAT) capture. As the Research Electrical Engineer responsible for Light Stage hardware, Bipin Kishore watched the resolution and fidelity targets of the Deep3DMM database and newer production scan work outgrow what those cameras could deliver. The problem was easy to state and harder to solve: the lab needed a single camera that could hit cinema-grade dynamic range and SNR while also meeting the synchronization, form factor, and linearity demands of the Light Stage.

The Challenge

From years of building and deploying Light Stage systems for both clients and lab research, Bipin had a clear sense of what the camera actually had to do. Unlike a typical studio or broadcast camera, anything mounted in the Light Stage has to satisfy all of the following at once:

1. Synchronize with precision-timed LED illumination pulses via Genlock and Timecode. Any frame-level misalignment corrupts the OLAT data.

2. Run at frame rates high enough to limit motion blur from the subject. Even small movements during a sequence introduce geometric inconsistencies across the multi-view array that hurt 3D reconstruction.

3. Stay highly linear across nearly the full dynamic range, so pixel values represent actual reflectance instead of sensor artifacts.

4. Hold high SNR at the exposure and ISO levels used during facial capture, so noise doesn’t bleed into the 3D reconstruction.

5. Be small enough to mount densely around the geodesic dome, and stay reliable there.

6. Allow custom firmware for Light Stage Master triggering, multi-camera control, and direct data offload to the server.

None of the existing cameras hit all of those constraints together. The Canon 1DX, capable as it was, didn’t have the frame rate the heavier research captures needed. The Ximea cameras had a nice mix of small form factor, frame rate, and global shutter, but couldn’t deliver the resolution or dynamic range needed for pore-level facial geometry. A pipeline-wide camera change wasn’t something Bipin was going to make on impressions alone, so he designed a proper evaluation to see whether the RED KOMODO could close all the gaps in a single body.

Objective

Primary Research Question

Can the RED KOMODO meet the photometric linearity, dynamic range, SNR, and synchronization requirements of the USC ICT Light Stage MVS facial capture pipeline, and is it actually a meaningful upgrade over the Canon 1DX and Ximea systems currently in use?

Bipin designed and ran the full evaluation himself. He’s the lead behind the design, deployment, and operation of every Light Stage hardware system at the lab, and has built and maintained seven complete Light Stage systems for clients including EA, Activision, and Meta. That background shaped how the evaluation was structured. The goal was to test the cameras against real production conditions, not just isolated bench numbers.

System Setup and Configuration

For the evaluation, he used two RED KOMODOs and wrote custom ICT UI and microcontroller firmware to lock the KOMODO’s Timecode across the Light Stage array. This wasn’t optional. Out-of-the-box Timecode would have let the cameras drift at the frame level and corrupted the data. He paired each camera with a Canon 85mm f/2.8 RF lens and a KOMODO Link adapter for the control and data access needed during synchronized multi-camera capture.

Camera settings for the evaluation:

Evaluation Configuration

• Resolution / Frame rate: 6K 17:9 @ 30FPS (maximum framerate @ High setting)

• ISO: 500 | Recording quality: HQ | Aperture: f/11

• Sync method: Genlock + Timecode via Light Stage Master (custom ICT system)

• Trigger: GPI from Light Stage Master PCB; frame rate sampled via GPO

• Output format: .R3D → OpenEXR (no compression) via REDline.exe

• Reference target: 90%-reflective white square of a standard color chart

Exposure values tested (multiples of 2):

1/30 — 1/60 — 1/125 — 1/500 — 1/1000 — 1/2000 — 1/4000 — 1/8000

Evaluation Methodology

Bipin built four experiments to characterize the KOMODO across the conditions a Light Stage actually puts a camera through. He used the same color chart and controlled illumination in all four so the results would line up cleanly. The only thing he changed between experiments was the LED intensity, which let him simulate the range of capture conditions the pipeline has to handle.

Experiment 1: SNR Under Flat Illumination (Linearity Baseline)

For the first experiment, he set the stage to Analog Intensity 100 (constant current at 20% of max) with PWM cranked to maximum to kill any LED flicker. The point was to get a clean baseline for sensor linearity under stable illumination, before pushing the camera into harder territory. He captured at least four frames per exposure setting and pulled pixel values from the 90%-reflective white patch of the color chart. The .R3D files were converted to uncompressed OpenEXR so nothing in the analysis pipeline would alter the signal.

Then, going further than a simple white patch, he sampled pixel values across the full greyscale strip of the color chart, plotting each one against the chart’s known reflectivity. That gave him a direct number for photometric linearity instead of just a visual check, which matters for any camera being used in reflectance modelling.

Experiment 2: SNR Under Underexposed Conditions

For this experiment, the Stage Analog Intensity was dropped to 25 (constant current at 5% of max) to simulate the low end of the exposure range. This shows up in certain OLAT sequences where individual LED brightness is minimal. The goal was to expose any non-linearity or signal convergence in the sensor’s response at low signal levels, since that’s where a camera that looks fine in normal conditions can quietly fall apart.

Experiment 3: SNR Under Overexposed Conditions

For the third experiment, the camera was pushed the other way, bumping Stage Analog Intensity to 200 (constant current at 40% of max) to drive the KOMODO into the top of its dynamic range. He wanted to find the exact point where the sensor’s response starts to break down: the threshold past which pixel values stop being trustworthy as reflectance measurements. That gives a clear upper boundary to exclude from production captures.

Experiment 4: Recording Quality Comparison — HQ vs MQ vs LQ

With the stage back at the standard Analog Intensity 100, he tested all three RED recording quality settings (HQ, MQ, and LQ) at the same time. He wanted to know whether the compression trade-offs in MQ and LQ would degrade SNR enough to hurt downstream 3D reconstruction. Whatever came out of this would shape the official capture protocol for different production scenarios, especially long performance captures where storage budgets get tight.

Results

Linearity and SNR

On all three channels, the RED KOMODO came out with much higher SNR than either the Canon 1DX or the Ximea cameras the lab had used before. The greyscale linearity tests showed it stays linear through the mid-tone range, which is the part that matters most for facial reflectance modelling.

Exposure Range Boundaries

The underexposed and overexposed runs gave Bipin two operational boundaries that went straight into the lab’s capture protocol:

• Lower boundary: the bottom 10% of the underexposed range shows the 2nd and 3rd exposure points converging, which means you can’t reliably tell them apart. This range is excluded from all production captures.

• Upper boundary: the top 10–15% of the overexposed range shows non-linear SNR increase and the curve structure breaks down. This range is also excluded from all production captures.

Recording Quality

MQ recording came surprisingly close to HQ at both ends of the range, so Bipin marked it as an acceptable fallback for long performance captures where storage gets tight. LQ didn’t hold up in the mid-to-upper range, so he ruled it out for any research-grade or production scan work.

Key Findings Summary

• Linear photometric response confirmed across the mid-tone range that matters most for facial reflectance capture

• Bottom 10% of the underexposed range excluded from the protocol because the exposure points converge there

• Top 10–15% of the overexposed range excluded because the SNR curve goes non-linear there

• HQ recording is now the institutional standard; MQ is an acceptable fallback for long performance captures

• LQ recording is excluded from all research-grade and production scan protocols

Camera Comparison: RED KOMODO vs. Previous Systems

The table below summarizes the key specs of the three camera systems evaluated. Figures marked [X] still need to be filled in from technical datasheets.

Going Live

Once the evaluation was done, Bipin led the integration of the RED KOMODOs into the USC ICT Light Stage capture pipeline. The migration touched almost every layer of the stack. He wrote custom firmware to lock Timecode across the multi-camera array, adapted the Light Stage Master control software for the KOMODO’s GPI/GPO triggering, redesigned the camera mounting hardware so the KOMODO body sat cleanly on the geodesic dome, and built a new R3D-to-EXR conversion workflow around REDline.exe for the lab’s standard data pipeline.

What’s Next

The KOMODO-upgraded Light Stage pipeline has since produced facial scans used in visual effects work on major motion pictures and the work is far from done. Bipin is now evaluating the RED V-RAPTOR [X] for the next round of Light Stage builds that need even higher resolution and speeds. He's integrating the KOMODO pipeline into the lab's real-time ReFA (Rapid Face Asset Acquisition) workflows, and the array keeps going out to more than the 3 already client-deployed Light Stage systems. Next up: RED Connect for IP-based multi-camera streaming, which should strip most of the cabling complexity out of production scan setups.

About Bipin Kishore

Bipin Kishore is a Research Electrical Engineer at the USC Institute for Creative Technologies (USC ICT). He has designed and deployed seven Light Stage systems for EA, Activision, Meta, and major motion picture productions. He is a co-author on peer-reviewed publications in digital human capture and holds a U.S. patent. His photometric capture infrastructure is the hardware behind the ICT 3D Morphable Face Model, which has been commercially adopted by NVIDIA, Zoom, and Flawless.

Evaluation designed and conducted by Bipin Kishore, USC ICT — bipin.kishore45@gmail.com