This repo includes code to perform, and above all, visualize, object detection on:

• a) a cam feed (your webcam) :: *_cam_detect.py

• b) video feed (some video saved on your disk) :: *_video_detect.py

and so far supports two detection model formats:

• i) Tensorflow frozen graph (protobuf) :: tf_*
• ii) Tensorflow Lite (only with cam feed for now) :: tflite_*
• iii) Coral (yet to be supported)

Currently supported implementations are indicated by the *.py scripts at the root of this repo.

You need to have these files depending on what you want to run.

The practice established in this repo is that you save these files to the ./tf_files folder. A prototype whale/dolphin detection model is available here so you can get started. Place these files into the ./tf_filesfolder and adjust the paths in the scripts if need be. You can also pass these paths via command line, as per the following example --help output. All root ./*.py scripts have this option available.

$ python tf_cam_detect.py --help
usage: tf_cam_detect.py [-h] [-m GRAPH_PB] [-c GRAPH_CONFIG] [-p MIN_PROB]

optional arguments:
  -h, --help            show this help message and exit
  -m GRAPH_PB, --graph_pb GRAPH_PB
                        path to Tensorflow protobuf object detection model
  -c GRAPH_CONFIG, --graph_config GRAPH_CONFIG
                        path to pbtxt graph config file
  -p MIN_PROB, --min_prob MIN_PROB
                        minimum probability to filter detections

Besides all you can find on the webs, you can start with this wiki, written with this repo in mind. Also, this project started with this article from Tony607 (and its associated repo).