Last week I shared how vision models felt like they could “think.”
But the real insight wasn’t the model.
It was the friction.
Here’s how that friction led me to architect a system — and why I call it CoreframeAI 🧵
Thinking of a self loop agent which helps to clean dataset and relabel back as incremental training. How to handle such false positives semantics of classes in the relabel back process 🤔💭?
@NotebookLM Been using it for q/a session, would be better if the podcast would also reflect on the latest source given and more engagingly ask the user if the info they gave is understandable.
It’s like a free seminar but focus on the person. Anyways kudos! It’s still fun and informative.
Irl job update
finished extracting coco80 and filter only wanted classes. Since its all polygon type bounding box, asked claude on converting polygon to bounding box coordinate. And making sure it is not rotated 🤣. The fun begins when need to filter good quality images of 30k.
The amount of repeated labels and adjustments need to be done in this single image. For this case is for vehicle class. Lol 156/18796 after 200 i might stop for a while
Been trying out to test new MVP for VLM but stuck at runpod configs and not working scripts which are not tested in the first place lol. Burnt $1.24 runpod credits. Still worth it in my opinion.
#coreframeai#eisencfaivlm#buildinpublic