SparseEnd2End: Obstacle 3D Detection and Tracking Architecture Based VisionTransformer

---

| Blog |

👋 Hi, I’m ThomasVonWu. I'd like to introduce you to a simple and practical deployment repository based on TensorRT which uses end-to-end perception paradigm with sparse transformer to sense 3D obstacles. This repository has no complex dependency for Training | Inference | Deployment(which means, we don't need to install MMDetection3d, mmcv, mmcv-full, mmdeploy, etc.), so it's easy to install in your local workstation or supercomputing gpu clusters. This repository will also provide x86(NVIDIA RTX Series GPU) | ARM(NVIDIA ORIN) deployment solutions. Finally, you can deploy your e2e model onborad through this repository happily.
👀 I guess you are interested in:

• how to define a PyTorch custom operator: DeformableAttentionAggr and register related ONNX node.
• how to build a custom opertator plugin: DeformableAttentionAggr in TensorRT engine with Makefile or CMake.
• how to convert ONNX format file with custom opertator to TensorRT engine and make it as part of the whole engine.
• how to validate inference results consistency : PyTorch results vs. ONNX Runtime results vs. TensorRT results.
• how to convert PyTorch model with temporal fusion transformer head to ONNX.
• how to locate the TensorRT layer accurately when overflow occurs during using fp16 quantization for model parameter.

Algorithm Architecture

Algorithm Framework of Sparse4D, which conforms to an encoder-decoder structure. The inputs mainly consists of three components: multi-view images, newly initialized instances, propagated instances from previous frame. The output is the refined instances (3D anchor boxes and corresponding features), serve as the perception results for the current frame. Additionally, a subset of these refined instances is selected and propagated to the next frame.

Experimental Results

Model	ImgSize	Backbone	Framework	Precision	mAP	NDS	FPS	GPU	config	ckpt	onnx	engine
Sparse4Dv3	256x704	Resnet50	PyTorch	FP32	56.37	70.97	19.8	NVIDIA GeForce RTX 3090	config	ckpt	--	--
Sparse4Dv3	256x704	Resnet50	TensorRT	FP32	wait	wait	wait	NVIDIA GeForce RTX 3090	config	ckpt	onnx	engine
Sparse4Dv3	256x704	Resnet50	TensorRT	FP16	wait	wait	wait	NVIDIA GeForce RTX 3090	config	ckpt	wait	wait
Sparse4Dv3	256x704	Resnet50	TensorRT	INT8+FP16	wait	wait	wait	NVIDIA GeForce RTX 3090	config	ckpt	wait	wait
Sparse4Dv3	256x704	Resnet50	TensorRT	FP32	wait	wait	wait	NVIDIA ORIN	config	ckpt	wait	wait
Sparse4Dv3	256x704	Resnet50	TensorRT	FP16	wait	wait	wait	NVIDIA ORIN	config	ckpt	wait	wait
Sparse4Dv3	256x704	Resnet50	TensorRT	INT8+FP16	wait	wait	wait	NVIDIA ORIN	config	ckpt	wait	wait

News

• 24 Sep, 2024: I release repository: SparseEnd2End. The complete deployment solution was released.
• 25 Aug, 2024: I release repository: SparseEnd2End. The complete deployment solution will be released as soon as possible. Please stay tuned!

Tasklist

• Register custom operation : DeformableAttentionAggr and export ONNX and TensorRT engine. 25 Aug, 2024
• Verify the consistency of reasoning results : DeformableAttentionAggr PyToch Implementation vs. TensorRT plugin Implementation. 25 Aug, 2024
• Export SparseTransFormer Backbone ONNX&TensorRT engine. 8 Sep, 2024
• Verify the consistency of reasoning results : SparseTransFormer Backbone PyTorch Implementation vs. ONNX Runtime vs. TensorRT engine. 8 Sep, 2024
• Export SparseTransFormer Head ONNX and TensorRT engine. 24 Sep, 2024
• Verify the consistency of reasoning results : SparseTransFormer Head PyTorch Implementation vs. TensorRT engine. 24 Sep, 2024
• Reasoning acceleration using CUDA shared memory and CUDA FP16 in DeformableAttentionAggr plugin Implementation.
• Reasoning acceleration using FlashAttention in replace of MultiheadAttention.
• Reasoning acceleration using FP16/INT8 in replace of FP32 of TensorRT engine.
• Reasoning acceleration : Image pre-processing Instancbank Caching and model post-processing Implementation with CUDA.
• Image pre-processing, Instancbank Caching and model post-processing Implementation with C++.
• Onboard: Full-link reasoning using CUDA, TensorRT and C++.

Introduction

SparseEnd2End is a Sparse-Centric paradigm for end-to-end autonomous driving perception.

Quick Start

Quick Start

Citation

If you find SparseEnd2End useful in your research or applications, please consider giving me a star 🌟

🏷 ChangeLog

08/25/2024： [v1.0.0] This repository now supports Training | Inference in NuscenesDataset. It includes: data dump in JSON, Training | Inference log caching, TensorBoard hooking, and so on.