rattydave/ai-object-detect repository overview

⁠Object Detection

The purpose is to get the object detection and proof of concept working in the minimum time.

Ethos:

• We use pre-compiled binaries where possible.
• The python code contains the minimal needed to be functional.

⁠Defaults

• Image size of 640x480
• ssdlite_mobilenet_v2_coco_2018_05_09

This uses pretrained models and can has the ability to change the model easy using the configuration file.

⁠Scripts

⁠To install run the following.

xhost +
docker run -it \
           --name camera_go \
           -e DISPLAY=$DISPLAY \
           -v /tmp/.X11-unix:/tmp/.X11-unix \
           --device /dev/video0 \
           rattydave/ai-object-detect:latest

⁠To start the object detetection run the following

xhost +
docker start -i camera_go

⁠Auto Update

To automatically update I recomend using watchtower.

docker run -d \
    --name watchtower \
    -v /var/run/docker.sock:/var/run/docker.sock \
    containrrr/watchtower

⁠To change the object detetection model

• start up the container with docker exec -it camera_go bash
• edit /root/download.sh and uncomment the download line.
• edit ~/tensorflow1/models/research/object_detection/obj-config.ini to change to the right model.
• run /root/startup.sh to start the detection.

⁠Models

All models are taken from Tensorflow detection model zoo⁠

The default is ssdlite_mobilenet_v2_coco_2018_05_09.

⁠COCO-trained models from COCO dataset⁠

90 Classes

⁠Kitti-trained models from Kitti dataset⁠

2 Classes

⁠Open Images-trained models from Open Images dataset⁠

601 Classes

⁠iNaturalist Species-trained models from iNaturalist Species Detection Dataset⁠

2854 Classs

⁠AVA v2.1 trained models from AVA v2.1 dataset⁠

AVA is a project that provides audiovisual annotations of video for improving our understanding of human activity.

90 Classes