Fine-Tune a Semantic Segmentation Model with a Custom Dataset
Hugging Face Blog
A guide demonstrates how to fine-tune SegFormer, a semantic segmentation model, to recognize sidewalks and obstacles for a pizza delivery robot using a custom sidewalk dataset. The training uses the smallest SegFormer variant (B0), which is only 14MB in size, enabling the model to run on the robot's hardware. After fine-tuning on 50 epochs with a learning rate of 0.00006, the model can be pushed to Hugging Face Hub and used for inference to segment sidewalk images in real time.
