hous-lab/sparsedrive-trt-edge

SparseDriveV2 TensorRT deployment ONNX

The four deployment ONNX exported by the sparsedrive-trt-edge deployment pipeline (this HF repo is the companion of the code repo, same name). Exported 2026-10-20 from the official SparseDriveV2-Small (B2D stage2) checkpoint (sha256 f15c889b759734894e6ac906403df55a9973d7e60e3101da3ac71e9091f7b130).

Files

File	Size	What it is
sparsedrive_backbone.onnx	107.3 MB	shared CNN backbone (ResNet-34 + FPN), raw export - no If-folding applies, this is the deployment ONNX
sparsedrive_head_sim.onnx	81.3 MB	3D detection head, If-folded deployment export (the _sim suffix)
sparsedrive_map_sim.onnx	108.3 MB	online mapping head, If-folded deployment export
sparsedrive_motion_plan_sim.onnx	68.3 MB	motion prediction + planning head, If-folded deployment export

The _sim suffix marks the If-folded deployment export: the training graph contains If subgraphs that TensorRT 10.3 cannot parse unfolded (see the code repo docs/lessons.md 7.16), so the pipeline folds them before building engines. Backbone has no conditional subgraphs and needs no folding.

Usage

The pipeline normally exports these itself (run_pipeline.sh step 2); downloading from here skips the export step - place the ONNX under onnx_export/output/ and continue from the engine-build step.

Build TRT engines with trtexec plus the repo's DeformableAggregation plugin:

trtexec --onnx=sparsedrive_motion_plan_sim.onnx --fp16 --staticPlugins=<path to libplugins built per the code repo>

TensorRT version pairing matters: x86 uses TRT 10.4; Orin/JetPack must use the JetPack-matched TRT 10.3 - 10.4 on Orin is measurably slower on the plugin-bearing graphs (docs/trt_versions.md in the code repo).

Full I/O contracts, accuracy gates and latency numbers: the code repo (docs/pipeline_guide.md, docs/engine_profiling.md).

Provenance and attribution

Exported from the official SparseDriveV2-Small (B2D stage2) checkpoint. Weights are NOT rehosted here - download from the upstream HuggingFace: https://huggingface.co/wenchaosun/SparseDriveV2 (upstream code MIT). This repo hosts only our own pipeline export artifacts, released under Apache-2.0 (pipeline: Apache-2.0).

Citation

If these artifacts help your work, please cite/acknowledge the upstream SparseDriveV2 work: https://github.com/swc-17/SparseDriveV2