Object Detection best model / best algorithm in 2023 | YOLO vs SSD vs Faster-RCNN comparison Python

In this video, we are going to see which is the best object detection algorithm or model for developers. We are going to test all the model based on three criterias: speed, accuracy and ease of implementation.

Object Detection is one of the most sought after domain of computer vision and the number of models available in this domain reflect the same, however not all models were create the same. Each model that we discuss in this video has its own pros and cons, but we are after those criterias that matters the most to us. We compared a model from the Two shot detector family which is Faster RCNN. The comparison also included two single shot models also, which are SSD (Single Shot Detectors) and YOLO. When comparing for speed, we focused on the inference speed of the models, ie how many frames can the model process in one second. For accuracy, we tried to see which model actually got the most accuracy and how reliable those accuracies are. Lastly, we also looked at the ease of implementation which basically focused on the framework (Opencv, PyTorch, TensorFlow) required to use the model and also the least number of lines of code we need to write to get the model to give detections.

[flueepwrien]

You could put the videos you are mentioning in the description for ease of use.

[sedatcakici]

Hello, just wanted to say that I love your content very much, it's very interesting and informative.
Thanks a lot

[SalihFCanpolat]

I would beg to differ, ease of implementation is the WORST for YOLOv8 if you wish to implement it yourself from ground zero. Let us say you wish to prune the model to be used along with security cameras or in an environment where the GPU does not support as many operations then you need to custom make your own architecture. It is easy to follow YOLOv1 for or R-CNN but for modern YOLOv8 the whole process is convoluted mess.

I believe data science channels should focus more on real world applications then data science bootcap lies.

[HighlyTheoretical]

What about when using a framework like open-mmlab

[floatonArt]

Bro can you please provide training script for the ssd and faster rcnn please

[phantomgaming]

What is the fastest and lightest model we can use? That detects/classifies object that are easy to detect. It should just have negligible impact on realtime detection

[subhankhan]

I disagree with the ease of implementation, it all depends on your own transfer learning or inference file. There are heaps of examples where you just parse items and use single CLI for both SSD and Faster-RCNN. Ultimate winner should be SSD due to it’s inference and latency plus the ability to be used commercially. Most of yolo algorithms are either AGPL or GPL licensed, which requires heaps of money to be used commercially!

[StudyEnablers]

So yolo v8 is winner, whats about detectron2.

[s]

Dear Sir, would it be possible to have private contact related to the above topic?