Skip to content

autoware_tensorrt_yolox#

Purpose#

This package detects target objects e.g., cars, trucks, bicycles, and pedestrians and segment target objects such as cars, trucks, buses and pedestrian, building, vegetation, road, sidewalk on a image based on YOLOX model with multi-header structure.

Inner-workings / Algorithms#

Cite#

Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, Jian Sun, "YOLOX: Exceeding YOLO Series in 2021", arXiv preprint arXiv:2107.08430, 2021 [ref]

Inputs / Outputs#

Input#

Name Type Description
in/image sensor_msgs/Image The input image

Output#

Name Type Description
out/objects tier4_perception_msgs/DetectedObjectsWithFeature The detected objects with 2D bounding boxes
out/image sensor_msgs/Image The image with 2D bounding boxes for visualization
out/mask sensor_msgs/Image The semantic segmentation mask
out/color_mask sensor_msgs/Image The colorized image of semantic segmentation mask for visualization

Parameters#

Name Type Description Default Range
model_path string Path to onnx model. \((var data_path)/tensorrt_yolox/\)(var model_name).onnx N/A
label_path string Path to label file. $(var data_path)/tensorrt_yolox/label.txt N/A
score_threshold float A threshold value of existence probability score, all of objects with score less than this threshold are ignored. 0.35 ≥0.0
≤1.0
nms_threshold float A threshold value of NMS. 0.7 ≥0.0
≤1.0
precision string Operation precision to be used on inference. Valid value is one of: [fp32, fp16, int8]. int8 N/A
calibration_algorithm string Calibration algorithm to be used for quantization when precision==int8. Valid value is one of: [Entropy, (Legacy Percentile), MinMax]. Entropy
dla_core_id float If positive ID value is specified, the node assign inference task to the DLA core. -1 N/A
quantize_first_layer boolean If true, set the operating precision for the first (input) layer to be fp16. This option is valid only when precision==int8. False N/A
quantize_last_layer boolean If true, set the operating precision for the last (output) layer to be fp16. This option is valid only when precision==int8. False N/A
profile_per_layer boolean If true, profiler function will be enabled. Since the profile function may affect execution speed, it is recommended to set this flag true only for development purpose. False N/A
clip_value float If positive value is specified, the value of each layer output will be clipped between [0.0, clip_value]. This option is valid only when precision==int8 and used to manually specify the dynamic range instead of using any calibration. 6.0 N/A
preprocess_on_gpu boolean If true, pre-processing is performed on GPU. True N/A
gpu_id integer GPU ID for selecting CUDA device 0 N/A
calibration_image_list_path string Path to a file which contains path to images. Those images will be used for int8 quantization. N/A
Name Type Description Default Range
----------------------------- --------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------- ---------------
model_path string Path to onnx model. \((var data_path)/tensorrt_yolox/\)(var model_name).onnx N/A
label_path string Path to label file. $(var data_path)/tensorrt_yolox/label.txt N/A
score_threshold float A threshold value of existence probability score, all of objects with score less than this threshold are ignored. 0.35 ≥0.0
≤1.0
nms_threshold float A threshold value of NMS. 0.7 ≥0.0
≤1.0
precision string Operation precision to be used on inference. Valid value is one of: [fp32, fp16, int8]. fp16 N/A
calibration_algorithm string Calibration algorithm to be used for quantization when precision==int8. Valid value is one of: [Entropy, (Legacy Percentile), MinMax]. MinMax
dla_core_id float If positive ID value is specified, the node assign inference task to the DLA core. -1 N/A
quantize_first_layer boolean If true, set the operating precision for the first (input) layer to be fp16. This option is valid only when precision==int8. False N/A
quantize_last_layer boolean If true, set the operating precision for the last (output) layer to be fp16. This option is valid only when precision==int8. False N/A
profile_per_layer boolean If true, profiler function will be enabled. Since the profile function may affect execution speed, it is recommended to set this flag true only for development purpose. False N/A
clip_value float If positive value is specified, the value of each layer output will be clipped between [0.0, clip_value]. This option is valid only when precision==int8 and used to manually specify the dynamic range instead of using any calibration. 0.0 N/A
preprocess_on_gpu boolean If true, pre-processing is performed on GPU. True N/A
gpu_id integer GPU ID for selecting CUDA device 0 N/A
calibration_image_list_path string Path to a file which contains path to images. Those images will be used for int8 quantization. N/A

Assumptions / Known limits#

The label contained in detected 2D bounding boxes (i.e., out/objects) will be either one of the followings:

  • CAR
  • PEDESTRIAN ("PERSON" will also be categorized as "PEDESTRIAN")
  • BUS
  • TRUCK
  • BICYCLE
  • MOTORCYCLE

If other labels (case insensitive) are contained in the file specified via the label_file parameter, those are labeled as UNKNOWN, while detected rectangles are drawn in the visualization result (out/image).

The semantic segmentation mask is a gray image whose each pixel is index of one following class:

index semantic name
0 road
1 building
2 wall
3 obstacle
4 traffic_light
5 traffic_sign
6 person
7 vehicle
8 bike
9 road
10 sidewalk
11 roadPaint
12 curbstone
13 crosswalk_others
14 vegetation
15 sky

Onnx model#

A sample model (named yolox-tiny.onnx) is downloaded by ansible script on env preparation stage, if not, please, follow Manual downloading of artifacts. To accelerate Non-maximum-suppression (NMS), which is one of the common post-process after object detection inference, EfficientNMS_TRT module is attached after the ordinal YOLOX (tiny) network. The EfficientNMS_TRT module contains fixed values for score_threshold and nms_threshold in it, hence these parameters are ignored when users specify ONNX models including this module.

This package accepts both EfficientNMS_TRT attached ONNXs and models published from the official YOLOX repository (we referred to them as "plain" models).

In addition to yolox-tiny.onnx, a custom model named yolox-sPlus-opt-pseudoV2-T4-960x960-T4-seg16cls is either available. This model is multi-header structure model which is based on YOLOX-s and tuned to perform more accurate detection with almost comparable execution speed with yolox-tiny. To get better results with this model, users are recommended to use some specific running arguments such as precision:=int8, calibration_algorithm:=Entropy, clip_value:=6.0. Users can refer launch/yolox_sPlus_opt.launch.xml to see how this model can be used. Beside detection result, this model also output image semantic segmentation result for pointcloud filtering purpose.

All models are automatically converted to TensorRT format. These converted files will be saved in the same directory as specified ONNX files with .engine filename extension and reused from the next run. The conversion process may take a while (typically 10 to 20 minutes) and the inference process is blocked until complete the conversion, so it will take some time until detection results are published (even until appearing in the topic list) on the first run

Package acceptable model generation#

To convert users' own model that saved in PyTorch's pth format into ONNX, users can exploit the converter offered by the official repository. For the convenience, only procedures are described below. Please refer the official document for more detail.

For plain models#

  1. Install dependency

    git clone git@github.com:Megvii-BaseDetection/YOLOX.git
cd YOLOX
python3 setup.py develop --user

  2. Convert pth into ONNX

    python3 tools/export_onnx.py \
  --output-name YOUR_YOLOX.onnx \
  -f YOUR_YOLOX.py \
  -c YOUR_YOLOX.pth

For EfficientNMS_TRT embedded models#

  1. Install dependency

    git clone git@github.com:Megvii-BaseDetection/YOLOX.git
cd YOLOX
python3 setup.py develop --user
pip3 install git+ssh://git@github.com/wep21/yolox_onnx_modifier.git --user

  2. Convert pth into ONNX

    python3 tools/export_onnx.py \
  --output-name YOUR_YOLOX.onnx \
  -f YOUR_YOLOX.py \
  -c YOUR_YOLOX.pth
  --decode_in_inference

  3. Embed EfficientNMS_TRT to the end of YOLOX

    yolox_onnx_modifier YOUR_YOLOX.onnx -o YOUR_YOLOX_WITH_NMS.onnx

Label file#

A sample label file (named label.txt) and semantic segmentation color map file (name semseg_color_map.csv) are also downloaded automatically during env preparation process (NOTE: This file is incompatible with models that output labels for the COCO dataset (e.g., models from the official YOLOX repository)).

This file represents the correspondence between class index (integer outputted from YOLOX network) and class label (strings making understanding easier). This package maps class IDs (incremented from 0) with labels according to the order in this file.

Reference repositories#