This week you'll learn about An Incremental Improvement with Darknet-53 and Multi-Scale Predictions (YOLOv3).

Image

YOLOv2 was considered (at its time) the peak of object detection algorithms. It was faster than its competitors, could be run across various datasets, and the input size could also be altered. 

However, several competitors were popping up. At that time, since the goal of reaching a stable and fast object detector was thought to have been achieved, it was all about making it more accurate. 

This led to the birth of YOLOv3, which is widely considered an INCREMENTAL IMPROVEMENT over YOLOv2. Today we learn about the intricacies of YOLOv3, another impactful addition to the YOLO family. 

The big picture: YOLOv3 brought in many notable changes of its own. These additions helped propel the YOLO family to the top of the object detection game again. 

How it works: YOLOv3 utilized several concepts like confidence-based bounding box predictions and multilabel predictions. It also improved its performance in detecting small objects, where the previous YOLO versions struggled. YOLOv3 also utilized Darknet-53, a hybrid predecessor of Darknet-19 (used by YOLOv2). A unique feature of YOLOv3 is its ability to predict bounding boxes at different scales throughout its architecture. 

Our thoughts: Performance-wise, YOLOv3 showed significant improvements in many areas over other detection systems. 

Yes, but: It was seen that one crucial metric showed YOLOv3 decreased in performance, specifically that it couldn't get the bounding boxes perfectly aligned with the required object. This meant that even though YOLOv3 was faster than the other methods, there was scope for more work. 

Stay smart: The YOLO family doesn't end here. So, stay tuned for our next entry in this series! 

Click here to read the full tutorial

Solve Your CV/DL Problem This Week (or weekend) with our Working Code

You can instantly access all of the code for An Incremental Improvement with Darknet-53 and Multi-Scale Predictions (YOLOv3) by joining PyImageSearch University. Get working code to

  1. Finish your project this weekend with our code
  2. Solve your thorniest coding problems at work this week to show off your expertise
  3. Publish groundbreaking research without multiple tries at coding the hard parts

Guaranteed Results: If you haven't accomplished your CV/DL goals, let us know within 30 days and get a full refund.

I want the code

Note: You may have missed this, but last Wednesday, we published a new post on Planning Next Steps When Things Don't Work Out.

The PyImageSearch Team

Image