jetson_tensorrt

Implemented Nodes

DIGITS ImageNet (classification)

• classify_nodes.launch runs a classifier trained on ILSVRC2012 with the builtin camera and publishes to /rt_debug

Parameters

Param	Type	Description
image_subscribe_topic	string	image topic to run classification on
model_path	string	absolute path to the model file (.prototxt)
weights_path	string	absolute path to the weights file (.caffemodel)
cache_path	string	absolute path to the automatically generated tensorcache file
classes_path	string	newline delimited list of class descriptions starting at id 0
data_type	int	TensorRT data type. 32 for kFLOAT, 16 for kHALF, 8 for kINT8
model_image_depth	int	model input image depth / number of channels
model_image_width	int	model input width in pixels
model_image_height	int	model input height in pixels
threshold	float	confidence threshold of classifications, between 0.0 and 1.0
mean1, mean2, mean3	float	ImageNet means

Topics

Action	Topic	Type
publish	classifications	Classifications
subscribe	image_subscribe_topic	Image

Messages

# Classification
uint32 id
float32 confidence
string desc

# Classifications
ClassifiedRegionOfInterest[] regions
Header header

DIGITS DetectNet (detection)

• detect_nodes.launch runs pedestrian detection on the builtin camera and publishes to /rt_debug

Parameters

Param	Type	Description
image_subscribe_topic	string	image topic to run detections on
model_path	string	absolute path to the model file (.prototxt)
weights_path	string	absolute path to the weights file (.caffemodel)
cache_path	string	absolute path to the automatically generated tensorcache file
classes_path	string	newline delimited list of class descriptions starting at id 0
data_type	int	TensorRT data type. 32 for kFLOAT, 16 for kHALF, 8 for kINT8
model_image_depth	int	model input image depth / number of channels
model_image_width	int	model input width in pixels
model_image_height	int	model input height in pixels
model_stride	int	model stride size - this determines size of network outputs
threshold	float	confidence threshold of detections, between 0.0 and 1.0
mean1, mean2, mean3	float	ImageNet means

Topics

Action	Topic	Type
publish	detections	ClassifiedRegionsOfInterest
subscribe	image_subscribe_topic	Image

Messages

# ClassifiedRegionOfInterest
int32 x
int32 y
int32 w
int32 h
uint32 id
float32 confidence
string desc

# ClassifiedRegionsOfInterest
ClassifiedRegionOfInterest[] regions
Header header

Planned Nodes

• DIGITS - SegNet

Requirements

• Jetpack 3.3
• TensorRT 4.0

Build / Installation

Clone jetson_tensorrt into your catkin_ws/src folder and run catkin_make

Documentation

• Doxygen