Object detection with Owl-ViT

This project implements the Owl-ViT model for zero-shot object detection in videos (or several images).

Detection using the prompts "person" and "ball".

Detection using the prompts "person" and "balloon".

Contents

Table of contents:

• Installation
• Usage
• Example

Installation

Run the following commands:

git clone https://github.com/killian31/ObjectsDetection.git
cd ObjectsDetection
pip install -r requirements.txt --upgrade

Usage

Usage: 
    python3 owl_vit.py [-h] --imgs_dir IMGS_DIR --save_to SAVE_TO [--process_video] [--video_filename VIDEO_FILENAME] [--image_start IMAGE_START] [--image_end IMAGE_END]
                       [--texts TEXTS [TEXTS ...]] [--thresholds THRESHOLDS [THRESHOLDS ...]] [--box_thickness BOX_THICKNESS] [--save_model]

Options:
  -h, --help            show this help message and exit
  --imgs_dir IMGS_DIR   The directory containing the images.
  --save_to SAVE_TO     the directory in which to save the processed images
  --fps FPS             Number of frames per second for the output video (default: None, determined automatically if --process_video).
  --process_video       Wether to get images from a video or not (default: False).
  --video_filename VIDEO_FILENAME
                        Name of video file to process (default: None).
  --image_start IMAGE_START
                        Frame to start from (default:0).
  --image_end IMAGE_END
                        Frame to end with (default: last (0).
  --texts TEXTS [TEXTS ...]
                        A list of texts to detect in the images (default: 14 random texts).
  --thresholds THRESHOLDS [THRESHOLDS ...]
                        A list of thresholds between 0 and 1 (default: 14 low thresholds for the texts).
  --box_thickness BOX_THICKNESS
                        The thickness of the bounding boxes to draw (default: 2).
  --save_model          Whether to save the pretrained model locally or not (default: False).

Run python3 owl_vit.py -h to display the help above message.

Example

python3 owl_vit.py --imgs_dir frames --save_to detected_example --process_video --video_filename data/video.mp4 --texts person ball --thresholds 0.08 0.12 --save_model