Apple’s Depth Pro 2 estimates meters from a single camera. What should we measure?
Depth Pro 2 handles images and video. Distance error, boundary accuracy and stability across frames matter for different reasons.

Apple’s research team has made the code and weights for Depth Pro 2 available. Its arXiv submission is dated November 19, and by November 20 Apple’s model page listed four sizes ranging from about 24 million to 872 million parameters. The family estimates pixel-level depth in meters from an ordinary camera image, a video or a live feed at the input resolution. The weights are downloadable under a license for non-commercial scientific research and academic development, not a general commercial SDK or a released consumer feature.
Ordering a scene is not the same as measuring it
A visual effect that places an animated character between a coffee table and a sofa needs plausible occlusion. It might look right whether the table is one meter from the camera or two. A room-planning tool is different: an estimate off by twenty centimeters can affect decisions about furniture placement. Both tasks use depth, but the first leans on ordering and edges, while the second depends on whether the estimated scale survives a check against known distances.
An ordinary photograph is ambiguous about scale. A small nearby object and a larger, farther object can occupy the same number of pixels. Models use information about the camera and patterns learned from data to make an estimate. Apple provides a focal-length estimator for cases where that input is unknown. It helps fill a gap; it does not add a new physical measurement to the original image. Changing the lens or cropping the image can alter the cues a model sees, so success on one camera is not evidence of the same accuracy across devices.
Sharp boundaries and steady video solve different problems
A blurred depth edge can blend a nearby object with a distant background. A vase outline only a few pixels off can show a strip of background where an effect should have hidden it. In video, individually crisp estimates can still jump from one frame to the next as the camera moves. Heavy smoothing can cause the opposite problem: the depth map lags after an object moves. A consistent error remains an error, and a good still image does not establish stable motion. Apple’s paper aims to address those trade-offs with one network, but its reported comparisons are a starting point for testing, not proof for a particular application.
A useful evaluation keeps these outcomes apart. Compare absolute depth to known distances. Measure offsets separately at object boundaries. Run both static and moving clips to check jitter and lag. For a narrow object, distinguish between an error of a few pixels and one of several centimeters in the implied scene. A small set of independently measured indoor scenes can be re-captured with different lenses, lighting and devices. Break out reflective or transparent surfaces, thin objects and overlapping edges so a good global mean does not hide a recurrent failure mode.
Size the model against the complete workflow
The four sizes make an accuracy and speed comparison possible, but weight size is not end-to-end latency. Camera capture, preprocessing, moving data and drawing a result all take time. Average frame rate can also conceal intermittent pauses. Record the displayed latency distribution, dropped frames and how long predictions take to settle after a quick change. An on-demand estimate can trade speed for a sharper mask; a live preview may be far less tolerant of an occasional slow frame.
Depth Pro 2 makes these experiments easier to run with one camera. It does not turn a statistical estimate into an instrument reading. For compositing, prioritize occlusion and visual continuity. For decisions that depend on dimensions, compare estimates against distances measured independently and document the devices and conditions under which the result is reliable enough to use.
Sources
Information only. No investment, legal, tax, or financial advice.