search

utilForever / CubbyDNN

Deep learning framework using C++17 in a single header file
MIT License
30 stars 6 forks source link

Implement Image class #27

Open Yoogeonhui opened 5 years ago

Yoogeonhui commented 5 years ago

To put image to a tensor, implementation of image class is needed

Basic Features

Additional features

Ravenwater commented 5 years ago

How does this design work together with video streams? One very important use case of a DL environment is object recognition and tracking, with our without optical flow.

Make certain that the image class works well together with video processing workloads.

Yoogeonhui commented 5 years ago

There'll be a video class which can return an image class per frame. And the video class will be implemented after this work is completed.

Ravenwater commented 5 years ago

Awesome.

What about optical flow-based video analysis? Do you know what is needed for that use case? I don't know, so asking in earnest.

Yoogeonhui commented 5 years ago

I'm not familiar with flow-based video analysis. I saw a dataset whose r, g and b represents x, y, z-axis movements respectively. But actually I don't know what is needed for that use case specifically.

Ravenwater commented 5 years ago

We should look into what the best practice is for optical flow. I believe that most insect visual systems are optical flow based, and I would think that this translates to DL for autonomous systems such as drones. It might be a separate cognitive system, so a fancy drone may have a 3D visual system for object recognition and tracking, and an optical flow DL system for control.

Let's make certain that CubbyDNN can support these use cases.

utilForever commented 5 years ago

We separate task that implements load images.