Object Detection Lab
In lab:
- Use MobileNets and separable depthwise convolutions.
- Use 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.
Requirements
Install environment with Anaconda:
conda env create -f environment.yml
Change TensorFlow pip installation from tensorflow-gpu
to tensorflow
if you don't have a GPU available.
The environment should be listed via conda info --envs
:
# conda environments:
#
carnd-advdl-odlab /usr/local/anaconda3/envs/carnd-advdl-odlab
root * /usr/local/anaconda3
Further documentation on working with Anaconda environments.
Particularly useful sections:
https://conda.io/docs/using/envs.html#change-environments-activate-deactivate https://conda.io/docs/using/envs.html#remove-an-environment
Resources
- TensorFlow object detection model zoo
- Driving video
Tips
- Some users have reported the driving video as playable only in Jupyter Notebook operating in Chrome browser, and not in media player or Jupyter Notebook operating in other browsers. In contrast the post-segmentation video appears to be operating accross players and browsers.