This is a Model using ResNet-50 backbone and customized Multi-task Head and loss to make classification, boundary box prediction and segmentation (24.7M parameters).

Note that I only use processor of detr-resnet-50-panotic and resnet-50 backbone of the base model, not use transformers. All the model, task heads and loss are self-defined.

Another model by directly fine-tuning DETR model can be found at https://huggingface.co/FriedParrot/fish-segmentation-simple

This model use kaggle dataset A Large Scale Fish Dataset as dataset for training. And for convenience, I also made a copy version for this dataset available on huggingface, this is just for making it easier for u to use.

Tasks :

Classification
BBoxes prediction, and
Segmentation

[!warning] Since this model include customized type, then AutoModel() and AutoConfig() may fail, but AutoProcessor() will work correctly (Since I use a DetrImageProcessor for this)

If you want using this model, you can go to my source code below and look for FishSegmentationModel and FishSegmentModelConfig for load these models correctly.

Model Sources

For source code & Tutorials : check my github

Results and test

I trained this model in my pc(RTX4060 8GB + cu126), and those are some pictures tested in fish datase :

(The model predicted the mullet a shrip lol😂 since classification head of this model is not very accurate😂)

FriedParrot/fish-segmentation-model

Fish-segmentation-model

Model Sources

Results and test

FriedParrot/
fish-segmentation-model