FriedParrot/fish-segmentation-simple

Model Card for Fish Segmentation (Fine-Tuned DETR)

This is a fine-tuned DETR model (`facebook/detr-resnet-50-panoptic`) adapted for fish detection and segmentation. The model performs multi-task prediction including:

• —Classification (fish species recognition)
• —Bounding Box prediction
• —Segmentation masks

It has 42.9M parameters and is trained on the [A Large Scale Fish Dataset](https://www.kaggle.com/datasets/crowww/a-large-scale-fish-dataset) from Kaggle.

The copy of this dataset on hugging face is available here

Model Sources

• —Base model: facebook/detr-resnet-50-panoptic
• —Fine-tuned model: FriedParrot/fish-segmentation-simple
• —Training dataset: A Large Scale Fish Dataset
• —Source code & tutorials: GitHub Repository

[!note] This model is fully compatible with AutoModelForObjectDetection, AutoProcessor, and Hugging Face Trainer. Unlike the first model (fish-segmentation-model), this one does not require custom config classes.

Training Details

• —Hardware: NVIDIA RTX 4090 (48GB VRAM)
• —CUDA: 12.8
• —Framework: PyTorch + Hugging Face Transformers
• —Batch size: use 8 as train batch sizes
• —Training strategy: Direct fine-tuning of DETR with minimal modifications

Results & Example Predictions

Since its a fine-tuned model, the accuracy is really high, and also classification accuracy can reach about 100%.

The predicted bounding box and masks are also very accurate :