PeopleNet tlt model based on DetectNet v2

These models accept 960x544x3 dimension input tensors and outputs 60x34x12 bbox coordinate tensor and 60x34x3 class confidence tensor.

url : https://ngc.nvidia.com/catalog/models/nvidia:tlt_peoplenet

In this code we aimed to inference from a tensorrt model which is obtained from a tlt model and test the outputs.

To this end, you need to convert your tlt model into trt by using two commands:

1. tlt-export to convert your tlt model into etlt
2. tlt-converter to convert your etlt model into trt

multiple batch

url : https://github.com/NVIDIA/DL4AGX/blob/master/MultiDeviceInferencePipeline/training/objectDetection/ssdConvertUFF/utils/inference.py

Requirements

tlt container (version >= 3) which supports TensorRT and pycuda

1. tensorrt version : 7.2.1.6
2. cuda : cuda-11.1

An etlt model: tlt-export detectnet_v2 -m model_input.tlt
-k tlt_key
-o model_output.etlt
-e spec_files/train_spec_file.txt
--batch_size 1
--data_type fp32

A trt model: tlt-converter -k tlt_key
-d channel,height,width
-o NMS
-e model_output.trt Your trt model name
-m 1
-t fp16
-i nchw
model_input.etlt Your etlt model name

Running the code

for running the model, one can easily run the python3 example.py the container:

python example.py

Parameters in the example file:

1. input of the example code is an image.
2. the output of the example code is a list of dictionaries which includes 3 options: class_id, confidence and bounding_box

[{'class_id': 2, 'confidence': 1.0, 'bounding_box': [0.629, 0.13, 0.088, 0.169]}

class_id values are like this:

class_id	category
0	person
1	bag
2	face

the format of bounding box output is like this(percentage of width_image and heigh_image): [x_min, y_min, w_bbox, h_bbox]