GTR-Depth
A depth-estimation model in the GTR series, built on a gated linear attention architecture. It predicts metric absolute depth from monocular images without relying on softmax attention, and supports real-time inference at millisecond-level latency.
Timeline (last 90 days)
GTR-Depth can predict absolute metric depth in meters from a single monocular image.
GTR-Depth is a continuation of the GTR series of models, focused on the depth estimation task.
GTR-Depth uses the same backbone network as its predecessor, containing 12 gated linear attention blocks.
GTR-Depth uses no softmax attention anywhere in the entire network.
GTR-Depth achieves a single-frame inference latency of 2.7 milliseconds on the NVIDIA DRIVE AGX Thor platform.
GTR-Depth has not yet disclosed the specific implementation of its scale constraint.
Publicly available information about GTR-Depth comes primarily from a Reddit post and does not yet include comprehensive accuracy benchmarking data or training dataset details.
A latency of 2.7ms theoretically implies the ability to process more than 300 frames per second.
All Facts (8)
GTR-Depth can predict absolute metric depth in meters from a single monocular image.
50%UnverifiedGTR-Depth is a continuation of the GTR series of models, focused on the depth estimation task.
50%UnverifiedGTR-Depth uses the same backbone network as its predecessor, containing 12 gated linear attention blocks.
50%UnverifiedGTR-Depth uses no softmax attention anywhere in the entire network.
50%UnverifiedGTR-Depth achieves a single-frame inference latency of 2.7 milliseconds on the NVIDIA DRIVE AGX Thor platform.
50%UnverifiedGTR-Depth has not yet disclosed the specific implementation of its scale constraint.
50%UnverifiedPublicly available information about GTR-Depth comes primarily from a Reddit post and does not yet include comprehensive accuracy benchmarking data or training dataset details.
50%UnverifiedA latency of 2.7ms theoretically implies the ability to process more than 300 frames per second.
50%