Right now is honestly a great time to get into 3D stuff with blender and threejs etc. They've been brewing in adolescence for like a decade but it's all suddenly really mature and exciting right now (and mobile GPUs have caught up nicely). If anyone finds themselves asking "where the heck are the low hanging fruit anymore?" (like I did recently) the answer is in 3D content for the web and other places. It's a much more exciting space than ML (which I played with before that), where things are mostly just in the hopeful research zone at the moment.
+1 on this. Years ago I looked at it a bit but the interface was only for people who knew what they are doing and my laptop was way too slow anyway.
This year, I’m giving it another try with the infamous Doughnut tutorial and I’m very impressed. Once I got a general idea on the thing, I’m able to navigate the interface and figure out how to do stuff without too much effort. It’s incredibly powerful piece of software(it has scripting and visual programming capabilities) with a very healthy community.
Oh an my Macbook Air M1 is more than enough to get it going. Disney doesn’t need to get scared just yet but having access to such a powerful software on such a portable device makes you feel powerful.
I agree. Very exciting time and feels much more accessible now than when I tried it a few years back. I'd also reccomend trying unreal engine 5 which is royalty free up to $1M in revenue or something.
You can also create in blender and then import into unreal.
The wonderful thing is that since Blender is totally scriptable in Python, and Python is the lingua franca of machine learning, you can seamlessly and efficiently use Blender to do all kinds of cool things with machine learning and AI!
Both creating ML-enhanced content creation and editing tools of course, but also synthesizing images, videos, 2D and 3D content for ML training.
For example, you can use using GPU-accelerated ray tracing and procedural character animation in photorealistic environments to train body tracking and pose estimation networks, since you get the precise "ground truth" and pixel-perfect semantic segmentation for free without enlisting Mechanical Turk to tediously and unreliably annotate millions of video frames.
And you can use randomization and retexturing to plug in different body types, skin tones, hair styles, clothing, etc, to train it to work with diverse sets of people, in a wide range of environments, camera angles, and lighting conditions.
>Using blender to generate realistic images for ML training:
>This is a two parts talk. It is based on the papers "3D Scene Reconstruction from a Single Viewport" presented at ECCV 2020 and the "BlenderProc" paper. The speaker is the main author of both papers. This is the recording of part 2.
>Lecture abstract: We present BlenderProc, which is a modular procedural pipeline, helping in generating real looking images for the training of convolutional neural networks. These can be used in a variety of use cases including segmentation, depth, normal and pose estimation and many others. A key feature of our extension of blender is the simple to use modular pipeline, which was designed to be easily extendable. By offering standard modules, which cover a variety of scenarios, we provide a starting point on which new modules can be created.
