Paper 1 , Paper 2 darkflow darknet
- Automatically trains a class via webscraping image search results on a video recognition classifier with transfer learning.
- Enter a label, then enter a list of search queries to google for. It will then google for those search terms and fine-tunes a pretrained classifier.
- Detect objects as well as output alerts if an object in your "alert list" is found.
- Can be performed on a video stream in real-time.
- Can be performed on a live-camera stream in real-time.
- Python 3
- ffmpeg
- OpenCV
- OpenCV-Python
- Tensorflow-GPU
- CUDNN and CUDA ToolKit
On the webscraper, indicate the label, as well as the search terms to use by editing it and change the parameters in downloadimages.py then in batch, type
python downloadimages.py
Once images are downloaded. You can download pretrained weights here: darknet Or you can continue training your weights if you've done this before Edit the parameters in train.py and then in batch type:
python train.py
You can test on either a video file, through video_run.py You can test on a live webcam feed through camera_run.py
python video_run.py
python camera_run.py
You will be asked for the path of the video file. You can adjust the parameters as well as the paths of the weights by opening up the py files.
You can test on either a video file, through video_run.py You can test on a live webcam beed through camera_run.py
python video_run.py
python camera_run.py
You will be asked for the path of the video file. You can adjust the parameters as well as the paths of the weights by opening up the py files.
You can set alerts by editing the text file "alerts.txt" when a label found in this text file appears, it will generate an alert by drawing the box red and displaying "Alert x found in footage" when testing.