[Revised Edition] Deep Learning Computer Vision: The Complete Guide

hello,

In the previous announcement, we informed you that there was a problem with Keras (not tf.keras, but the original Keras) when downgrading to TensorFlow 2.3 or lower in Google Colab.

Although this issue has been registered as a bug, Google Colab currently only leaves a message recommending setting up a tensorflow version (i.e. 2.4 or higher) that is compatible with the Keras version, and there has been no further progress.

Based on my personal experience, I expect that it will take longer than expected to resolve this error, and I believe that this will result in a gap in practical training for those who use Colab as a practice environment.

Accordingly, we have decided to sequentially release the revised version of this lecture in advance to minimize the gap in the practice environment.

The revised edition will consist of approximately 80% of the existing lectures, and will introduce additional lectures compared to the existing lectures. It mainly focuses on the following points.

1. More detailed explanation of the theory regarding the many questions students have had

2. Practice based on the latest/highest performing Object Detection/Segmentation package

3. Reflecting the latest trends in Object Detection/Segmentation

4. Writing more diverse, expandable and flexible practice codes and more detailed explanations.

5. Various other additional lectures

6. Focus on hands-on practice based on Google Colab environment due to difficulty in GPU allocation in Google Cloud

In order to move away from the package of this lecture, which has been implemented in Tensorflow 1.x, we have invested a lot of time and tested several packages. As a result of the test, we have selected OpenMMLab's MMDetection and Ultralytics' Yolo v3 packages as the packages for practice, and are currently in the process of creating a practical lecture.

As those who know know, MMDetection and Ultralytics' Yolo package are currently (as of May 2021) the most recognized packages in the field of Object Detection. MMDetection is also called Kaggle's winning solution and is recognized for its excellent performance while implementing various cutting-edge algorithms. Ultralytics' Yolo is yolo v5, which caused a stir in this world(?) last year. Although we will practice with yolov3, yolov5's interface is almost the same as v3, so yolo v5 can be applied without major changes.

One thing that is unfortunate is that both of these packages are based on pytorch. However, since the current lecture is aimed at those who have a basic understanding of tensorflow and keras, the revised lecture has little or no pytorch code. In order to connect with the current lecture and reduce the time required to produce the lecture, I invested a lot of time to test Object Detection/Segmentation packages based on Tensorflow 2.x, but I have to admit that they still fall short of the advantages of MMDetection and Ultralytics Yolo.

The original revised version was planned to be opened in mid-July. However, since there is an issue with the current lecture's colab practice environment, I plan to upload the completed practice lecture videos first, and I will probably upload 2-3 per day starting next Monday. And I will do my best to update all lectures to the revised version by the end of June. I can say with certainty that it is composed of better and more detailed practice lectures than the current lectures.

The detailed revised lecture update schedule will be announced again before next Monday. The revised version will increase the tuition, but of course, existing students will not be affected. I would like to thank the students who loved my lectures once again, and I will work on the schedule quickly to minimize the gap in practice.

If you have any questions regarding the sequential release of this revised edition, please feel free to leave a question.

thank you