Parallax
About

A lab for perception that keeps pace with the camera.

Parallax builds the models, runtime and tools that let cameras, robots, glasses and kiosks understand a scene in 8 ms, offline, on the device itself.

Mission

Make every camera able to respond.

Billions of cameras are already installed. Almost all of them only record.

For a machine to act on what it sees, perception has to finish before the next frame arrives. Streaming video to a data centre and waiting for an answer takes hundreds of milliseconds, and it depends on a network that is rarely there when it matters.

We think the answer is to move understanding to the device. That takes models built for a fixed budget, a runtime that schedules many of them inside one frame, and sensors chosen with the model in mind. Parallax works on all three.

8 msFrame to decision, on the device
0Frames sent to a cloud by default
4Layers built together, sensor to SDK
2Labs, Zurich and San Francisco
What we believe

Four ideas we build around.

They decide what we work on and, as often, what we leave alone.

Latency is a feature.

A machine that sees late acts late. We treat the frame budget as part of the specification, not a number to tune at the end.

Video should stay where it was filmed.

Processing on the device is faster, and it means footage of people and places never has to travel to be understood.

Models and hardware are one design.

The sensor, the ISP and the accelerator shape what a model can do. We design them together, from the pixel up.

Publish what moves the field.

We share methods, benchmarks and evaluation sets when they help others build real-time systems, and we say how we measured.

The lab

Research that has to run inside a frame.

Every result is measured on target hardware before we call it a result. These are the areas we spend most of our time on.

Depth and geometry

Monocular and stereo depth that runs in a few milliseconds, with calibrated confidence on every pixel.

Tracking and memory

Stable identities across occlusion and motion, with state that lives on the device between frames.

Sensor co-design

Choosing exposure, readout and ISP settings jointly with the model, so the camera captures what perception needs.

Where we are

Two labs, one bench.

Most of the team works in person, close to the devices we ship to.

Zurich

Research, compilers and hardware. Most of our model training and sensor work happens here, next to the bench.

San Francisco

Runtime, product and partner engineering. This is where teams bring their devices for a first deployment.

Backed by

People who have shipped hard systems.

We are funded to do long research and to ship it.

Parallax is backed by engineers and investors from robotics, semiconductor and developer-tools companies, people who have taken perception, silicon and SDKs from a prototype to millions of devices.

  • Founders and early engineers of robotics and drone companies
  • Architects of mobile and automotive neural accelerators
  • Builders of developer platforms used by large engineering teams

We are hiring in both labs →

Put real-time vision on your device.

Install the SDK, point it at a camera and see your first detections in an afternoon.