←
Deepen AI211 posts
Deepen AI avatar

Deepen AI

@deepenai

Safety, productivity, velocity!

◷ Joined January 2017▧ 52 Photos and videos
Deepen AI avatar
Deepen AI@deepenai·29 Jan 2020

Safety Pool™ brings together leading automotive industry organizations, policymakers, regulators and stakeholders to build a shared collaborative future for Safe Autonomous Driving and ADAS systems. Excited to have @Aptiv on board! #SafetyPool #AV #Safety #ADAS @wef

Post attachment
Deepen AI avatar
Deepen AI@deepenai·14h

A calibration error can misplace LiDAR points by about 50 cm at 35 m. A safety score measured on that data inherits the error. So PASAP characterizes the data first: calibration integrity, label quality, ODD coverage, and real-world data to fill the gaps.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 17

"Our data is high quality" is a marketing claim. Before you buy a dataset, you should be able to see samples, the price, and how it was recorded. Ask for the listing behind the claim. Refinery by Deepen shows samples and prices up front: refinery.deepen.ai/catalog

Deepen AI avatar
Deepen AI@deepenai·Nov 16

Industry standard for calibration reprojection error: 1.5 pixels. Deepen holds under 0.8, across 200+ customer deployments. The 0.7px gap is the difference between calibration that passes a procurement checkbox and calibration a perception engineer trusts.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 16

An invoice can reconcile perfectly and still be wrong in four places: an O read as a zero, a day and month swapped, a supplier name cut short, a currency that was only a default. Arithmetic checks miss all four. Reviewing a sample of the documents that passed catches them.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 13

A data entry pilot: two people, 25 hours a week, for two weeks, at $5 per person-hour. That's $500 to see how a trained team handles your records before you commit to anything longer.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 10

For manipulation, calibration error isn't abstract. A grasp is planned in 3D from camera geometry. If the geometry is off, the gripper arrives off target by that much. Sub-centimetre precision shows up directly in grasp success.

Deepen AI avatar
Deepen AI@deepenai·Nov 9

LiDAR to camera calibration is now self-serve. Upload a recording you already have (ROS1, ROS2 or MCAP), no targets needed, and get the extrinsics back in typically 15 to 25 minutes. $150 at launch pricing (list $1,500) until 31 Dec: targetless.deepen.ai

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 9

VLMs need better-grounded data, not just more of it. If calibration drift has moved the sensor's frame of reference, the caption and the pixels describe different scenes. The model learns that mismatch, and its errors lean the way the drift does.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 6

Well-calibrated data reduces labeling overhead by up to 30%. When calibration drifts, annotators stop labeling what objects are and start fixing where they are. That's data troubleshooting billed at labeling rates.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 5

Reprojection error, in one line: take a known 3D point, project it into the camera image using your calibration, and measure how far it lands from where it actually appears. That gap is geometry your perception model learns wrong.

Deepen AI avatar
Deepen AI@deepenai·Nov 5

A freelance marketplace and a managed team are different purchases. On a marketplace you choose and manage the people, and quality depends on who you picked. With a managed team you describe the task once and get a named lead, weekly QA and one published rate.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 3

How to record a drive for targetless calibration: 30 to 60 seconds, with a figure-8 if you have the space. No targets, no calibration bay. The motion gives the solver what it needs to align LiDAR, cameras, IMU and GNSS in one pass.

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 2

New categories come from what partners list and what buyers ask for. Got data in one of these? list@deepen.ai

Post attachment
Deepen AI avatar
Deepen AI@deepenai·Nov 2

Every listing uses the same format: samples you can evaluate, capture details, permitted uses, and validation status, as-is or through Deepen. Our own data follows the same rules.

Post attachment