README

Simple object detection library using mobilenet ssd arachitectures.

Supports

• Tensorflow cpu support
• Tensorflow gpu support
• TF-TRT support(nvidia tensorflow docker required)
• Native Tensorrt support(nvidia tensorflow docker required)(Tensorrt version 7+)
• Auto download object and face models
• Support for L4T containers

Pre requisite Requirements

• Tensorflow 1.X
• Opencv

Installation

pip3 install git+https://github.com/MacherLabs/objectDetectorSSD.git@tensorrt

How to use

#Import
import numpy as np
from objectDetectionSSD import ObjectDetectorSSD

#Prepare inputes
model_name='frozen_inference_graph.pb' #Frozen pb graph
trt_enable=True # TF-TRT conversion enablement
tensorrt_enable=False #Tensorrt conversion enablement
precision='FP16' # Precision to convert to
classes=['person','car'] # Classes to predict(if [], predicts all classes in labels.json)
thresh=0.5 # Threshold for detection confidence
input_shape=300 # Input shape of model
gpu_frac = #Amount of gpu to allocate, if 0 allows growth

#Initialize detector
detector=ObjectDetectorSSD(model_name=model_name,
                        gpu_frac=gpu_frac,
                        tf_trt_enable=trt_enable,
                        tensorrt_enable=tensorrt_enable,
                        precision=precision,
                        classes=classes,
                        threshold=thresh,
                        input_size=input_shape
                    )
# Generate test image
img =np.random.rand(300,300,3).astype('uint8')

# Perform detections
dets=detector.detect(img,threshold=thresh)

Benchmarking on jetson nano

Mobilenet_300_ssd- 17 fps
Mobilenet_512_ssd- 9 fps