Road Segmentation for Autoware YabLoc (yabloc_pose_initializer)

Road semantic segmentation model used by the yabloc_pose_initializer package in Autoware.

The camera_pose_initializer node estimates the vehicle's initial pose from a camera image at the request of AD API. It segments the road surface in the undistorted camera image with this model and matches the result against the Lanelet2 vector map to score initial pose candidates.

The model is Intel Open Model Zoo's road-segmentation-adas-0001, converted to a TensorFlow frozen graph and to ONNX via the PINTO model zoo (entry 136). The node loads the frozen graph on CPU through OpenCV DNN; the ONNX export is shipped alongside it as the migration path off OpenCV's legacy TensorFlow importer (see Why ONNX is included).

Repository versions

Tag Contents Purpose
v0.1 The complete upstream PINTO model zoo export, all 104 files Archival snapshot of exactly what Autoware's previous hosting served, kept so the provenance chain survives the decommissioning of the old S3 bucket
v1.0 saved_model/model_float32.pb and saved_model/model_float32.onnx What Autoware installs. main tracks this tag

Consumers must pin a revision explicitly, never main. Autoware's ansible artifacts role pins v0.1 while the artifact hosting migration lands, so that move delivers exactly the same files as the old hosting did, and switches to v1.0 in a separate follow-up (see autowarefoundation/autoware#7223).

v0.1 additionally contains the TFLite variants (float32, float16, dynamic-range and full integer quantized), the TensorFlow SavedModel with its checkpoint, the TensorFlow.js graph models, the Intel OpenVINO IR pairs (FP32 and FP16) and the Myriad blob, and two TF-TRT converted SavedModels with their prebuilt TensorRT engines. None of them is loaded by Autoware. The TF-TRT engines in particular are 2021-vintage TensorRT 8.0 plans that cannot deserialize on a current stack: Autoware builds TensorRT engines locally on first launch and never ships them. They are archived in v0.1 for completeness only.

Model overview

Task Road semantic segmentation of a camera image, used for camera-based initial pose estimation
Origin Intel Open Model Zoo road-segmentation-adas-0001, converted by the PINTO model zoo
Runtime OpenCV DNN (cv::dnn::readNet), OpenCV backend, CPU target
Formats TensorFlow frozen graph float32 (model_float32.pb), ONNX opset 11 (model_float32.onnx)
Network input 512 x 896 RGB image, float32, scale 1.0, no mean subtraction. The frozen graph presents it as a 1 x 3 x 512 x 896 NCHW blob through OpenCV; the ONNX graph boundary is NHWC, 1 x 512 x 896 x 3
Network output 4-channel softmax score map (background, road, curb, marking), NCHW for the frozen graph and NHWC for the ONNX
License Apache-2.0 (Intel Open Model Zoo)

For the class definitions and architecture details of the segmentation network, see the upstream Intel Open Model Zoo model page.

Files

Contents of v1.0:

File Description
saved_model/model_float32.pb TensorFlow frozen graph, float32; the file the node loads today
saved_model/model_float32.onnx The same network exported to ONNX opset 11 by tf2onnx; not loaded today, see below
deploy_metadata.yaml Deployment metadata recording the artifact version of this repository

The directory layout (saved_model/model_float32.pb) is preserved exactly as the package's launch file expects it.

Why ONNX is included

The frozen graph and the ONNX file are the same network with the same weights, but they do not behave the same on current OpenCV releases.

Running the node's own call sequence (cv::dnn::readNet, OpenCV backend, CPU target, blobFromImage at 896 x 512, forward) against model_float32.pb:

OpenCV Softmax normalized over Result
4.6.0, 4.8.1 channel axis correct, the 4 class scores sum to 1.0 per pixel
4.9.0, 4.10.0 width axis wrong, class scores sum to 0.0059 and the output mean is exactly 1/896

On 4.9 and newer, OpenCV's TensorFlow importer applies the terminal softmax along the wrong axis. The output shape and the layer name are unchanged and no error is raised, so the failure is silent: at the node's default score threshold of 0.5 no pixel is ever selected. model_float32.onnx produces correct output on all four versions.

Autoware currently builds against OpenCV 4.5.4 (Ubuntu 22.04) and 4.6.0 (Ubuntu 24.04), which are unaffected. Ubuntu 26.04 LTS ships OpenCV 4.10.0. The ONNX file is therefore distributed so the consuming node can move to cv::dnn::readNetFromONNX, ONNX Runtime or TensorRT without needing the deprecated hosting or a reconversion. Note that the ONNX graph is NHWC at its boundary, so the node's blob construction and output conversion need to be adapted when it switches.

Inputs and outputs (as used by the node)

Subscriptions

Topic Type Description
~/input/camera_info sensor_msgs/msg/CameraInfo undistorted camera info
~/input/image_raw sensor_msgs/msg/Image undistorted camera image
~/input/vector_map autoware_map_msgs/msg/LaneletMapBin vector map

Publications

Topic Type Description
~/debug/init_candidates visualization_msgs/msg/MarkerArray initial pose candidates (the package README lists this topic as output/candidates, but the node publishes it under debug/init_candidates)

Services

Service Type Description
~/yabloc_align_srv autoware_internal_localization_msgs/srv/PoseWithCovarianceStamped initial pose estimation request

Pre-processing and post-processing run in the node: the image is resized to 896 x 512 and converted to a float32 RGB blob; the 4-channel output score map is resized back to the image resolution, the first (background) channel is dropped, and the remaining three channels are thresholded into a binary mask image used for map matching.

The node's only ROS parameter besides model_path is angle_resolution (default 30, the number of divisions of the 1 sigma angle range).

Usage in Autoware

Autoware downloads this artifact to ~/autoware_data/ml_models/yabloc_pose_initializer/ during environment setup (the ansible artifacts role). To fetch it manually:

hf download AutowareFoundation/yabloc_pose_initializer --revision v1.0 \
  --local-dir ~/autoware_data/ml_models/yabloc_pose_initializer

The package's launch file resolves the model at:

$HOME/autoware_data/ml_models/yabloc_pose_initializer/saved_model/model_float32.pb

and passes it to the node as the model_path parameter (overridable via the model_path launch argument). The node is started as part of the YabLoc localization stack; see the package README for details. If the model is missing, initialization still completes, but accuracy may be compromised.

Training

This model was not trained by the Autoware project; it is redistributed as-is from upstream:

Training datasets, schedules, and metrics are not documented in the Autoware sources; refer to the Intel Open Model Zoo model page for upstream details.

Provenance and versioning

Original source https://autoware-files.s3.us-west-2.amazonaws.com/models/yabloc/136_road-segmentation-adas-0001/resources.tar.gz (unversioned tarball)
Upstream mirror https://s3.ap-northeast-2.wasabisys.com/pinto-model-zoo/136_road-segmentation-adas-0001/resources.tar.gz (PINTO model zoo entry 136)
Source resources.tar.gz sha256 1f660e15f95074bade32b1f80dbf618e9cee1f0b9f76d3f4671cb9be7f56eb3a (the archive itself is not redistributed here; v0.1 holds its extracted contents)
saved_model/model_float32.pb sha256 c4e373552f4efb91592ed99f49afc6df179f8a95174ceaddd1ab35692f435b47
saved_model/model_float32.onnx sha256 931f80db4faa9eb37fafaaec35cdd6c7498a0b8afb45e335a535da028d27e50a

Every file in v0.1 is byte-identical to the corresponding member of that tarball.

Limitations

  • Inference runs on CPU via OpenCV DNN; the node does not use a GPU for this model.
  • The network operates at a fixed 896 x 512 input resolution; images are resized by the node.
  • The model was trained upstream by Intel, not on Autoware-specific data; segmentation quality on cameras or scenes that differ from the upstream training domain is not characterized here.
  • Pose initialization quality depends on the vector map and the undistorted camera input; a missing or poorly matching segmentation degrades the initial pose accuracy.
  • The frozen graph is affected by the OpenCV importer regression described above on OpenCV 4.9 and newer.

References

Acknowledgment

Special thanks to openvinotoolkit/open_model_zoo and PINTO0309 for providing and converting the original model.

Legal Notice

The original model is distributed by Intel Open Model Zoo under the Apache License, Version 2.0. The PINTO model zoo conversion scripts are released under the MIT license. See the upstream repositories for the full license terms.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support