Test drive with YOLO v4
YOLO V4 is the latest single phase Object detection model Alexey Bochkovskiy et. al. Its just fast and accurate. In this blog we will have a practical taste for ourselves what it has to offer.
YOLO models have been the talk in the object detection community right from its first release. It was June of 2015, Joseph Redmon et. al submitted a paper titled You Only Look Once: Unified, Real-Time Object Detection which left the research community in awe. They successfully overcame the age-old bottleneck in object detection; Real-Time Object Detection. The underlying master idea was to treat the existing problem as regression rather than classification. A lot has happened since then. Alexey Bochkovskiy et al proposed the latest version, YOLO v4 in April 2020.
- You Only Look Once: Unified, Real-Time Object Detection, 2015.
- YOLO9000: Better, Faster, Stronger, 2016.
- YOLOv3: An Incremental Improvement, 2018.
- YOLOv4: Optimal Speed and Accuracy of Object Detection, 2020.
Joseph Redmon took the world by surprise with this tweet
Here are some results which I obtained to while playing with YOLO v4
The identification of chairs in the image speaks about robustness.
The predictions of bottles, cups, and vase in the presence to huge aspect ratio difference to persons is noteworthy.
This is video at normal frame rate recorded on my camera. We can see the co-ordinates change in real-time with the movement of the subject.
This video has a faster frame rate than the previous one. Am struck with the FPS(frames per second) at which YOLO v4 predicts and still retain the accuracy in predictions.
There is a lot of interesting theory and experimentation behind the current state of the model. It has been an iterative process over the years. YOLO v4 has a lighter version that can perform detection in realtime on mobile phones. You can read about the changes in architecture over the years on the official Github repo and joe’s website.
Coming straight to the point! The model works wonders!. So much can be achieved using this strategy. The applications are limited to our imaginations and training data. The model is available trained on MS COCO data set on 80 classes. One can easily custom train the model on any dataset and utilize this beautiful creation. I had been a fan of YOLO v3 and that YOLO v4 is out a month now, I will write how to custom train YOLO v4 for your on CV projects in another blog.
You can access the colab notebook from the link here and replicate the results on your data. The instructions are available on the official Github Repo also.
