mapluisch / CV-DetectBottleCaps

Small YOLOv4 based bottle cap detection script written in Python. Computer Vision lecture.

Cheers - "Detect The Bottle Caps"

Cheers is a tiny Python3 script, which

1. reads some input video file
2. detects and extracts a still frame
3. runs inference on the still frame
4. detects and annotates any matching bottle caps (face-down/face-up/deformed)

I've trained my model using YOLOv4-tiny using ~500 images for training (~200 originally turned into ~500 using image augmentation).

Sample

Here's a quick sample of both a detected and extracted still-frame and the annotated result image.

It takes Cheers ~492ms in total to generate both result images; the input was a test-video file by me (~11 seconds, 1080p, 7MB).

Still-Frame detection

The still-frame detection and extraction takes ~175ms.

Inference + Annotation

Inference plus image annotation takes ~160ms.

Verbose console output

Here's the verbose output you'll get, when running Cheers using the --verbose flag.

Dependencies

• OpenCV-Python
• Python3

Install / Run

Simply clone this repo or download + unzip. Then, simply run

python3 cheers.py --video some_vid.mp4

or

python3 cheers.py --image some_img.jpg

and you'll get an annotated still frame in ./uutput/annotated frames/ after approx. 0.2s for input images and approx. 0.5-1.0s for input videos.

Cheers can also be used to live-annotate a passed-in video. You can also specify your own confidence- and non-max-suppresion-thresholds; please take a look at the parameters below.

Arguments

### input ###
--video			path to video file (to first detect a still frame from and then run inference on)
--image			path to image file (to directly run inference on)

### live-inference and -annotation ###
--live			show live-annotation of the passed in video-file, frame-by-frame

### dnn-variables ###
--conf			confidence threshold [0.0-1.0] whether or not a detected object is a bottle cap
--nms			non-max-suppression threshold [0.0-1.0] used to select the best bounding box per detected object

### region of interest detection ###
--find-roi		find a region of interest to only detect bottle caps within the found roi
--show-roi 		display detected region of interest

### misc ###
--verbose		print extensive information per step during runtime
--show-result		immediately show the annotated output frame using cv2.imshow
--show-conf		display the confidence values in the annotated result
--save-stillframe	save the detected still frame (default = false)
--high-precision       	enables comparatively slow, but more precise still-frame-detection and inference