[ in action ]
27 fps on one TDA4VM, with YOLO alongside.
One TDA4VM and a 1 MP binocular camera. Stereo runs on the LDC and SDE accelerators, calibration and filtering on the C66x DSPs, and YOLO on the C7x. The GPU is unused.
Photo: the test setup, a TDA4VM board with the 1 MP camera. Video: the 1 MP camera with Hammerhead, shown in NODAR Viewer.
27 fps
At 1 MP, with YOLO running alongside
2048 x 1024
With a 2 MP camera, 10 fps, no downsampling
$103 hardware
$23 camera plus $80 TDA4VM board
[ hardware ]
Complete stereo depth for $103 of hardware.
$23
Binocular camera, 1 MP, 20 g
$80
TI TDA4VM board in production volumes
Off-the-shelf silicon and an off-the-shelf camera give embedded 3D perception a shorter path from prototype to production. Already have cameras? We can run Hammerhead on yours.
[ resource budget ]
Good stereo does not need a big GPU.
Measured on one TDA4VM running Hammerhead and YOLO together at 27 fps.
Rectification VPAC LDC
64%
Stereo matching DMPAC SDE
70%
Calibration + filtering C66x DSP 1
33%
Calibration + filtering C66x DSP 2
33%
Object detection (YOLO) C7x / MMA
22%
System and application Cortex-A72 cores
under 50%
GPU
Off
blue: hammerhead · orange: yolo example app · grey: system and app
That leaves room for more models on the C7x, about half of each A72 for tracking and control, and an untouched GPU.
Calibration in every frame.
Immediate recovery from shocks and drift.
Watch the first frame.
Then the second.
The stereo cameras start out misaligned, and the depth map of this Pelican case falls apart. By frame two, Hammerhead’s online calibration has corrected the alignment, and the 3D structure snaps into place.
Why calibrate in real time?
Vibration
Robot motion and contact shake the cameras and can shift their positions, briefly or for good.
Thermal drift
Heat changes camera mounts and stereo geometry.
Error grows with range
A sub-millimeter shift in angle or baseline causes depth error that grows with the square of distance.
Failures cascade
Uncalibrated stereo causes SLAM drift, false obstacles and failed grasps.
Zero downtime
Continuous calibration replaces manual, target-based recalibration in the field.
No housing, no window.
Calibration does the housing’s job.
Rugged stereo cameras rely on machined housings, optical windows, stiffening cages, sealed connectors and gaskets, each adding cost and grams.
A housing’s other job is holding the two cameras rigidly aligned. Hammerhead’s online calibration takes over that job, so a conformal coating is enough to weatherproof the bare $23 board.
Shown: the coated camera submerged in water, still streaming images to the laptop.
[ humanoid robot evaluation ]
$23 camera vs. a $419 depth camera
A humanoid robot evaluation, both cameras at 1280 x 720. The $419 camera uses active IR with a pattern projector and a 95 mm baseline, and computes depth on board. The $23 camera uses passive stereo, with Hammerhead computing depth on a host processor (TI TDA4). Ground truth came from a laser rangefinder.
Optics set the precision. Calibration sets the accuracy.
13 mm
Precision at 3 m, vs. 15 mm
+15 mm
Accuracy error at 3 m, vs. +61 mm
20 g
Camera weight, vs. 116 g
Precision matched, accuracy 4x better at 3 m.
At 0.5 m the $419 camera returns no depth.
On a humanoid, every gram on the head counts.
20 g instead of 116 g.
1:4
Every 1 g on the head acts like 4 g of dynamic load, from inverted-pendulum physics
+1 to 10%
Longer runtime per battery charge
Longer life
Less wear on the lower actuators, joints and frame
Lower top mass means smaller motors, less peak torque, and lighter structures, such as plastics in place of metals.
[ minimum range ]
Full depth at 0.5 m,
where the $419 camera is blind.
Hammerhead searches a wider disparity range, so it sees closer. Tested down to 0.1 m.
$419 active-IR camera (95 mm baseline): no depth on the board
$23 camera + Hammerhead: full dense reconstruction
[ scaling to 5 mp ]
With a 5 MP camera, the gap widens.
Flat surfaces stay flat.
A $138 5 MP binocular camera with Hammerhead at 2592 x 1944 (12 cm baseline, 64° field of view, 10 fps), against the same $419 active-IR camera at 1280 x 720 and 30 fps. Same scene, same moment.
$419 active-IR camera: walls and floor break into ridges and dunes
5 MP camera + Hammerhead: flat surfaces reconstruct flat
Up close: on the $419 camera (left), the “nodar” logos on the beanie and sock are not resolved and a flat floor turns into waves. With the 5 MP camera and Hammerhead (right), both logos are legible and the floor reconstructs flat.
A ChArUco board is a known flat target. At 5 m, surface deviation is 33.1 mm RMS for the $419 camera and 7.1 mm RMS for Hammerhead, which also returns 34,810 points on the board against 3,227.
Measure
$419 active-IR camera
5 MP camera + Hammerhead
Precision at 6 m
38 mm
8.3 mm
Largest accuracy error, 3 to 6 m
250 mm
6 mm
Flatness on a board at 5 m
33.1 mm RMS
7.1 mm RMS
[ comparison ]
Three paths to a depth map.
[ getting started ]
From intro call to production.
Evaluate on your own data first.
The Hammerhead SDK for TDA4 includes 12 months of updates and support.
What’s included
Stereo matcher mapped to the TDA4 LDC and SDE accelerators.
Online calibration every frame.
Quality assurance with 16 stereo health metrics per frame.
Camera support from $23 1 MP boards to 5 MP automotive cameras.
GridDetect occupancy and obstacle detection, optional.
The steps
Start with a call about your use case, cameras, range and frame-rate targets.
01
Evaluation: Hammerhead on a TDA4VM with a 1 MP camera, or we process your recorded stereo data.
02
Proof of concept: on your hardware and cameras, benchmarked against your current depth sensor.
03
Production license: per-unit pricing by volume.