Parallax
Depth models

Perception models sized for the chip, trained for the sensor.

Six models that cover what most machines need to see. Each one fits in on-chip memory in int8 and leaves room in the frame budget for the others.

The family

Six models, one graph.

Detectors produce the features that Depth-Track reuses, so adding tracking costs about a millisecond, not a second network.

Detection

Depth-S

Fast general detector for people, vehicles and goods. The default on fixed cameras.

Parameters
4.1 M
Int8 size
4.6 MB
Input
960 × 544
Hexagon
6.4 ms
Jetson Orin NX
3.8 ms
Apple A18 Pro
4.8 ms
mAP 50–95, COCO-val
41.2
Detection + masks

Depth-M

Instance masks at 60 fps and up, for picking, counting and zones with odd shapes.

Parameters
11.8 M
Int8 size
12.9 MB
Input
1280 × 736
Hexagon
14.6 ms
Jetson Orin NX
8.1 ms
Apple A18 Pro
11.2 ms
mask mAP 50–95, COCO-val
38.6
Multi-object tracking

Depth-Track

Keeps identities across frames from detector features, without a second network pass.

Parameters
0.9 M
Int8 size
1.1 MB
Input
detector features
Hexagon
1.2 ms
Jetson Orin NX
0.8 ms
Apple A18 Pro
0.9 ms
HOTA, MOT17-test
63.4
Body keypoints

Depth-Pose

17 keypoints per person for ergonomics, gestures and safety zones around machines.

Parameters
6.3 M
Int8 size
7.0 MB
Input
640 × 640
Hexagon
9.8 ms
Jetson Orin NX
5.5 ms
Apple A18 Pro
7.4 ms
AP, COCO-keypoints val
68.9
Metric depth

Depth-Stereo

Dense depth from a stereo pair, in metres, for robots that need to know how far, not just what.

Parameters
8.7 M
Int8 size
9.4 MB
Input
2 × 960 × 544
Hexagon
12.4 ms
Jetson Orin NX
6.9 ms
Apple A18 Pro
9.6 ms
abs rel error, in-house warehouse set
0.041
Motion from event cameras

Depth-Event

Reads event-camera streams directly, for motion in the dark and at shutter speeds RGB cannot reach.

Parameters
2.4 M
Int8 size
2.7 MB
Input
event voxel grid
Hexagon
4.1 ms
Jetson Orin NX
2.8 ms
Apple A18 Pro
3.2 ms
flow EPE, DSEC-test
0.79 px

Median frame-to-decision, int8, batch 1, sensor-native input. Every chip and p99 on the benchmarks page.

Supported chips

Compiled for five chip families.

The Runtime compiles each model into a schedule for the exact part. New parts in a supported family usually work on day one.

Apple Silicon
A18 Pro · Neural Engine. Int8 schedules compiled by the Runtime, the full Depth family.
Depth-S 4.8 ms
Qualcomm Hexagon
QCS8550 · Hexagon NPU. Int8 schedules compiled by the Runtime, the full Depth family.
Depth-S 6.4 ms
NVIDIA Jetson
Orin NX 16 GB · GPU + DLA. Int8 schedules compiled by the Runtime, the full Depth family.
Depth-S 3.8 ms
Hailo
Hailo-8 · Dataflow NPU. Int8 schedules compiled by the Runtime, 5 of 6 models.
Depth-S 5.3 ms
Ambarella
CV72S · CVflow NPU. Int8 schedules compiled by the Runtime, 5 of 6 models.
Depth-S 7.3 ms
How we train for the sensor

The camera is part of the model.

Public datasets are photographs: framed by a person, exposed well, compressed once. A warehouse camera sees rolling shutter, sodium light and noise that climbs every time the dock door closes. We train on what the deployed sensor actually produces.

Characterise
Read noise, response curve, lens distortion and shutter timing, measured per sensor module.
Simulate
Training images pass through that sensor model in the raw domain, before demosaicing.
Co-design
Architectures are searched with the int8 model in the loop, so fragile layers never ship.
Calibrate
A two-minute capture in Studio picks the variant matched to each camera. No labels needed.
Bring your own

Your model, our Runtime.

Custom networks compile with the same tools we use for Depth. Export to ONNX, then compile for the target chip.

# compile a custom detector for Hexagon with exposure-banded int8
$ parallax compile ./my-detector.onnx \
    --target hexagon --precision int8 \
    --calibrate exposure-bands=4 --frames ./calib/aisle-12
# → my-detector.pxm · 5.2 MB · 7.1 ms median on QCS8550
  • ONNX opset 17 and later, with a report of every operator that falls back to the CPU.
  • The same profiler and traces as the Depth models, in Studio or on the command line.
  • Signed model packages that load offline, with nothing downloaded at run time.

Pick a model, see it run in an afternoon.

Install the SDK, point it at a camera and see your first detections in an afternoon.