vision icon indicating copy to clipboard operation
vision copied to clipboard

Add SSDlite Mobilenetv4

Open subhankhan041 opened this issue 8 months ago • 4 comments

🚀 The feature, motivation and pitch

Excellent work on SSDlite Mobilenetv3 and mobilnetv4. I was wondering if you guys provide an SSDlite Mobilenetv4 version just like we had in the past with SSDlite mobilenetv3.

Describe the solution you'd like: It would be interesting to see what difference it will have on the performance of this newer hybird version of SSDlite mobilenet v4.

Describe alternatives you've considered It's just an idea which could improve the real-time object detection based on SSD.

Alternatives

Try Schedulefree for optimization of this proposed SSDlite mobilenetv4.

Additional context

No response

subhankhan041 avatar May 14 '25 00:05 subhankhan041

Moving to torchvision repo

malfet avatar May 14 '25 14:05 malfet

Hey @subhankhan041, thanks for you message. We just shared an update on what we’ll be prioritizing in TorchVision (https://github.com/pytorch/vision/issues/9036). We do not plan to actively implement and release new models for now. But I would be interested to better understand your needs here. Are you asking about releasing model weights or the code for a specific model architecture?

AntoineSimoulin avatar May 22 '25 19:05 AntoineSimoulin

@AntoineSimoulin Thank you for the response. My impression is that if we have SSDlite-Mobilenetv4 and train it like it's previous version SSDlite-mobilenetv3 then we might see much better real-time object detection inference. So, in short if we have model weights for it then it would be great. This would definately contribute to the object detection research and could potentially revive SSD again.

ghost avatar May 26 '25 01:05 ghost

Hey @subhankhan041, thanks for all the details. We sync internally about your proposition. Unfortunately, we do not plan to actively neither implement nor release new models for now (#9036). Although we understand the appeal of releasing 'SSDlite-Mobilenetv4', our current focus is on other aspects of the codebase. We appreciate your understanding!

AntoineSimoulin avatar Jun 02 '25 13:06 AntoineSimoulin