CAMERA DEPTH · TI TIDL · C7x / MMA

See the scene in depth.

Recorded TI host emulation

Depth Anything V2 Small, repaired for the J721E TIDL toolchain. Original RGB on the left; the model’s estimated relative depth on the right.

Download videoModel + compiled artifacts ↗Validation reports ↗TI upstream PR #113 ↗
804 / 817ONNX nodes offloaded · 14 TIDL subgraphs
1.46%Worst video relative RMSE versus CPU
19 frames19/19 video frames passed

Dark: farther · Bright: nearer. Colours are scaled per frame, not metres.

Recorded video playback is 2 FPS. This is x86 TI host emulation with CPU Softmax and one CPU ConvTranspose. Execution and frame rate on a physical C7x/MMA board have not been verified. Agreement with CPU measures port fidelity, not real-world distance accuracy.

Compare with the original CPU model

CPU reference only. The main video above contains recorded TI provider outputs.

Why depth matters

Recognizing an object tells a machine what it is. Depth adds geometry: where the obstacle is and how much space is available to move.

LiDAR + cameras

LiDAR measures range. Camera depth adds dense image-aligned scene structure. Combining them requires sensor calibration, time alignment, and scale alignment.

ADAS

Calibrated depth and tracking support obstacle position, following distance, collision-risk estimates, and free-space planning. Relative depth alone cannot establish braking distance.

Robotics

Metric depth supports obstacle avoidance, 3D maps, reach planning, and grasp geometry. A robot needs calibrated geometry and stable tracking before using estimates for control.

Camera estimates are different from ToF measurements.

This model estimates relative inverse depth from a single RGB image. It does not measure distance in metres like a ToF camera or LiDAR. Dense geometry is useful, but physical scale needs additional information such as stereo, sensor fusion, or calibration. The demo predicts each frame independently.

What was repaired

LayerNorm recognition, Q/K/V selection, attention numerics, and five align-corners Resize operations. The Resize operations now use exact interpolation expressed as MatMul for MMA. The deployed graph retains the original learned weights.

Sources: Depth Anything V2 Small · Depth/LiDAR mapping · Automotive perception