This is an attempt to test a Python version of @pjreddie's YOLO, based on the code at PyImageSearch. The code has been modified for more

The repo contains 2 main python files: yolo_image.py and yolo_video.py They are for running the YOLO algorithm on single image and video (webcam by default) respectively, using Python.

The cfg folder contains the YOLO 2 weights and configs, as well as their tiny versions.

python3 yolo_video.py
	[--input|-i <video_file> (default: webcam)] [--output|-o <video_file>] [--cfg|-g <cfg_folder>]
	[--confidence|-c <min_confidence> (default: 0.5)] [--threshold|-t <threshold_nms> (default: 0.3)]
	[--fast|-f] [-s|--scaleRatio <video_scale_down_factor_for_faster_detection>]
	[--yoloVersion|-y (2 | 3 | 3-tiny) (default: 2)]

python3 yolo_image.py
	[--input|-i <image_file>] [--output|-o <video_file>] [--cfg|-g <cfg_folder>]
	[--confidence|-c <min_confidence> (default: 0.5)] [--threshold|-t <threshold_nms> (default: 0.3)]
	[--yoloVersion|-y (2 | 3 | 3-tiny) (default: 3)]

As the output file from running yolo_video.py is very fast, I made a python script to also slow down the output:

python3 make_video_slower.py
	[--input|-i <path_to_input_video>]
	[--output|-o <path_to_output_video>]
	[--fps|-f <_FPS_no> (default 3.0)]

To extract a bunch of frames from a video, I have made a simple python script extract_frames_from_video.py:

python3 extract_frames_from_video.py
	[--input|-i <video_file>]
	[--output|-o <path_to_folder_to_save_batch_images>]
	[--startFrame|-s <frame_no> (default: 0)]
	[--endFrame|-e <frame_no> (default: end of video)]
	[--displayFrames|-d]

• OpenCV: 3.4.4

• dlib

• numpy

• argparse

I tested YOLO versions 2 and 3 (and their tiny versions) against the Oxford Town Center dataset (truncated):

Confidence threshold: 0.5
Non-max Suppression threshold: 0.3

YOLO VERSION						v2		v3		v2(tiny)						v3(tiny)

Avg. Processing Time/Frame [sec]	0.25	0.65	0.05							0.05

Approx. Accuracy (mAP)				~15%	~90%	~15% (many false detections)	<10% (many false detections, worse than v2 tiny)

Confidence threshold: 0.3
Non-max Suppression threshold: 0.3

YOLO VERSION						v2		v3		v2(tiny)						v3(tiny)

Avg. Processing Time/Frame [sec]	0.3		0.5		0.05							0.05

Approx. Accuracy (mAP)				~20%	~90%	~15% (many false detections)	<10% (many false detections, worse than v2 tiny)

Note that I do not have a proper benchmark to gauge the accuracy. It's done by human guest-imate.

But you can see the video results can be found in videos/thresh<confidence_threshhold>/pedestrians_yolo<version>.mp4

• ASUS Zenbook Laptop (i7-8550, quad core), Ubuntu 18.10
• Nvidia GeForce MX150 GPU (2GB memory), CUDA 10.1, CUDNN 7.5.0

1. In terms of speed, YOLOv2 is almost 3 times faster than YOLOv3 (~0.15 seconds/frame vs ~0.6 seconds/frame). The difference in accuracy, however, is better than YOLOv2 (see videos/pedestrains_yolo2 -vs- 3).

Of course, the YOLO2/3-tiny versions are the fastest, but also super inaccurate.

2. From the comments on PyImageSearch, it's uncertain if the latest versions of OpenCV or dlib have GPU support now, but I actually suspect that the answer is YES.

Running YOLO2 on darknet's repo (supposed to use CUDA/GPU) gave me images processed in around 0.15 seconds. When I ran it with no GPU support, it was almost 10s. This python code gave images processed in about 0.2 seconds, so I think that GPU is supported now. (Not sure tho)

3. For some strange reason, the results of the YOLOv2 detection with this program is vastly different from that of the one run in the official darknet repo.

Below are the predictions, both with the same weights, same config and same confidence threshold (0.5). Both predictions are from the same frame (extractedFrames/5.jpg)[extractedFrames/5.jpg]:

Python Prediction

Darknet Prediction