Lite Any Stereo V2 for RDK S100/S100P β zero-shot stereo disparity
Compiled BPU models (.hbm) for the D-Robotics RDK S100 / S100P, ready to
load β no ONNX export, no calibration, no hb_compile. Built and measured with
BCDL, a C++17 inference and media library
for the RDK S-series with Python bindings.
Upstream: Lite Any Stereo V2 (M and L)
Redistributable, including commercially. The licence chain was checked on the code, the pretrained weights it started from, and the data it was trained on β all three, because a permissive repository badge does not by itself say anything about the weights. See Licence.
Files
| file | what it is |
|---|---|
las2_m_crop2_nashm.hbm |
M, centre-crop-fit calibration, 640x480 β 38.3 MB |
las2_m_resize2_nashm.hbm |
M, resize-fit calibration, 640x480 β 38.8 MB |
las2_l_crop2_nashm.hbm |
L, centre-crop-fit calibration, 640x480 β 67.8 MB |
las2_l_resize2_nashm.hbm |
L, resize-fit calibration, 640x480 β 67.6 MB |
las2_m_crop_nashm.hbm, las2_m_int16_nashm.hbm |
retired β clip disparity, see below; kept only so older BCDL fetch scripts keep working |
Measured on an S100P
| build | latency | throughput | EPE | bad 1 px | bad 2 px |
|---|---|---|---|---|---|
| M crop2 | 13.85 ms | 72 FPS | 0.998 px | 14.2% | 6.0% |
| L crop2 | 22.90 ms | 44 FPS | 0.928 px | 11.0% | 5.0% |
| M resize2 | 13.87 ms | 72 FPS | 0.597 px | 8.5% | 3.4% |
| L resize2 | 22.93 ms | 44 FPS | 0.568 px | 6.7% | 2.9% |
Latency: hrt_model_exec perf, one thread, 200 frames.
Accuracy: against ground truth on ETH3D and Middlebury 2021 scenes that are not
in the calibration set β 33 for crop, 51 for resize; the sets differ, so compare
within a mode. L's gain is close-range: on Middlebury, crop bad-1px goes 19.2%
β 14.5%, while on ETH3D the two are level. BPU time only β CPU
pre/post-processing is on top and is listed per task in BCDL's
benchmark results.
Use it
conda install -c https://mirrors.ruis.ai/conda -c conda-forge bcdl
import bcdl
engine = bcdl.Engine("las2_m_crop2_nashm.hbm")
print(engine.input_shape(0), engine.output_shape(0))
Each task has a decoder in BCDL that turns those raw outputs into boxes, keypoints, masks, disparity or text β see the Python API (δΈζ).
What to know before deploying
Two calibration modes, and you must match the one you letterbox with. _crop2
was calibrated on centre-cropped pairs and _resize2 on resize-fitted ones; feed a
build the geometry it was not calibrated for and the disparity degrades quietly
rather than failing.
The retired builds clip disparity. An int16 output threshold is the largest
value the calibration set produced. las2_m_crop_nashm and las2_m_int16_nashm
were calibrated on scenes with no near field and clip at 131 px and
47.7 px respectively, out of the model's 0β192 β closer objects read as
farther away, with no error. The 2 builds add shifted copies of each pair
(shifting the right view left by k raises every disparity by k) and reach
180β187 px. On the held-out resize set, EPE went from 2.17 px to 0.60 px.
Use the 2 builds.
All-2D and feed-forward β no 3-D cost volume β which is exactly why it runs at 72 FPS here. Cost-volume stereo of comparable quality took ~40 s/frame on this part and was abandoned.
Licence
MIT on the code and the weights. One thing worth knowing rather than a term passed down: the release was distilled from FoundationStereo, whose NVIDIA licence is research-only. That is upstream's compliance question, but look at it before a commercial build.
BCDL itself is Apache-2.0 and is unrelated to these terms β it is a
general-purpose runtime that loads any .hbm. The licence above constrains
these weights and this compiled artefact.
The conversion recipe β ONNX export, calibration, hb_compile config and the
acceptance numbers β is public in
bcdl-model-zoo, so this build can
be reproduced or retargeted rather than taken on trust.