I’ve been diving into object detection models lately and I came across this ConvNeXt thing. I remember trying to build a small object detector for a personal project a while back and it was a real pain. So, I’m curious if ConvNeXt is actually any good for this kind of stuff. Has anyone tried using it for object detection? Is it faster? More accurate? Or just another fancy name?
I like how ConvNeXt blends the simplicity of conventional CNNs with the strength of modern architectures, resulting in great accuracy and faster inference times.
ConvNeXt might not be the best fit for every application, especially if you need a model tailored to niche or highly specialized tasks.
Do you know where the depthwise convolution problem is found? I absolutely agree with you; on one of my projects, I encountered a similar issue that caused memory problems and hindered training (Pytorch). But as of yet, I haven’t come across any reliable information on this topic.
Perhaps quantization into F16 rather than F32 could help with this.