The goals / steps of this exercise are the following:
- Learn about MobileNets and separable depthwise convolutions.
- The SSD (Single Shot Detection) architecture used for object detection
- Use pretrained TensorFlow object detection inference models to detect objects
- Use different architectures and weigh the tradeoffs.
- Apply an object detection pipeline to a video.
Udacity project: link
Install environment with Anaconda:
conda env create -f environment.ymlChange TensorFlow pip installation from tensorflow-gpu to tensorflow if you don't have a GPU available.
- TensorFlow object detection model zoo
- ssd_mobilenet_v1_coco_11_06_2017 (~180 Mb)
- rfcn_resnet101_coco_11_06_2017 (~720 Mb)
- faster_rcnn_inception_resnet_v2_atrous_coco_11_06_2017 (~750 Mb)
- Driving video
- Fixed traffic cam video (https://www.youtube.com/watch?v=rpbkCUbWVio)
Output videos:
- ssd_mobilenet_v1_coco_11_06_2017
- rfcn_resnet101_coco_11_06_2017
- faster_rcnn_inception_resnet_v2_atrous_coco_11_06_2017
The object detection timing is depending on your hardware. The detection time roughly compares to each other:
ssd_mobilnet : rfcn_resnet101 : faster_rcnn_inception_resnet
80 ms : 320 ms : 1510 msIt is hard to find a good KPI to describe the detection quality. Model performance in a few word:
- The ssd_mobilnet usually don't find multiple car on a frame. One car is not detected on the lawn. Detection is not continuous on every frame when a car moves. Cars are not detected in the distance.
- The rfcn_resnet101 finds almost every car at least once. Detection is not continuous on every frame when a car moves. Cars are detected in crossroad.
- The faster_rcnn_inception_resnet finds almost every car at least once. Sometimes it even detects cars which are partially covered. Detection is almost continuous. Cars are detected in the crossroad.