Intel® OEP Sensor Fusion For Traffic Management - Release Notes#

Version 2026.1.0#

June 17, 2026

This release delivers BEVFusion 3D object detection enablement and optimization on Intel GPUs. It provides a complete end-to-end pipeline — from training and model export to optimized deployment inference — targeting autonomous driving and roadside perception (V2X) scenarios with multi-sensor (Camera + LiDAR) fusion.

Supported Platforms:

Platform

GPU

Intel Panther Lake (PTL)

Integrated GPU

Intel Arc B580 (Battlemage)

Discrete GPU

New

  • Sparse Convolution OpenVINO GPU Plugin Implementation

    Native 3D Sparse Convolution support in the OpenVINO GPU plugin, enabling the Second-based BEVFusion unified pipeline to run entirely within a single OpenVINO inference call on Intel GPU. The SparseConvolution operator (registered under domain org.openvinotoolkit) covers both SparseConv3d and SubMConv3d variants with fused BatchNorm + optional ReLU, totaling ~21 layers in the lidar sparse encoder. A custom OpenVINO build patch (custom_openvino_2026.1.0_sparse_ops.patch, ~12K lines) integrates all GPU kernel implementations into the OpenVINO 2026.1.0 GPU plugin.

  • BEVFusion-specific custom operators in GPU plugin

    SparseToDense (sparse feature map to dense BEV tensor conversion) and BevPoolV2 (camera-to-BEV view transform using precomputed geometry) are also implemented to support the full unified pipeline.

  • Two deployment pipelines

    • Split (PointPillars): ./bevfusion — 4 independent ONNX sub-graphs (camera backbone, lidar PFE, fuser, detection head) + external SYCL kernels, using standard ONNX / OpenVINO IR.

    • Unified (Second): ./bevfusion_unified — single unified ONNX with custom sparse ops executed inside the OpenVINO GPU plugin.

  • Multi-dataset support

    DAIR-V2X-I (V2X roadside) and KITTI-360, with geometry auto-detection from ONNX attributes via --preset v2x|kitti switch.

  • Training and model export toolchain

    Complete training-to-deploy workflow including dense mode training, BEVPool V1/V2 support, automated ONNX export, static-V PFE export, INT8 PTQ quantization (NNCF-based), and NVIDIA checkpoint compatibility (direct conversion from CUDA-V2XFusion .pth to Intel GPU deploy without retraining).

Improved

  • INT8 and FP16 inference optimization

    INT8 quantization via NNCF PTQ for both pipelines; FP16 inference mode for accuracy-first scenarios; mixed-precision support with per-stage INT8 toggles in the split pipeline.

  • SYCL-based high-performance kernels

    Hand-written SYCL kernels for voxelization (PointPillars and Second styles), BEV pooling, pillar scatter, and CenterHead post-processing (heatmap top-k, box decode, rotate-NMS).

  • Docker-based deployment

    Published Docker image intel/tfcc:2026.1.0-ubuntu24 with one-command smoke test and interactive container mode.

  • Visualization

    Built-in visualization interface with BEV and camera-view overlays (--save-image, --save-video, --display).

Known Issues

  • On Battlemage GPUs (Arc B580), the split pipeline falls back to FP16 for the fuser stage (fuser.onnx) due to a known INT8 fuser issue; other stages remain INT8.

  • Bundled release model assets use dummy weights for runtime interface validation; real trained weights are required for meaningful detection results.

On this page

This Page