pixel-level-segmentation-in-pets-unipr

Project objective:

• Pixel level segmentation in PETS dataset

Dataset:

• Pets dataset provides the segmentation mask of a large variety of pets https://www.robots.ox.ac.uk/~vgg/data/pets/

Network model:

• A segmentation model should be used for this task. UNet is a good choice, but you can experiments with different architectures.

Detailed information:

• Starting from a rgb image of a pet the network should output a segmentation mask of the pet. Only a single pet is present in each image.
• Intersection over Union, L1 distance are good metrics to evaluate the results.

Additional notes:

• Experiment also with in-the-wild samples (meaning images not extracted from the pets dataset: e.g. found online or taken by a smartphone). Does the network perform well over these images?
• That happens if more than on animal is in one image? How does the model perform?

---

Nets I decided to train/finetune and test

• UNet (required)
• DeepLabV3_ResNet101

Best performances achieved

	Model	Description	Resolution	IoU	L1 distance	Download
	UNet	Trained from scratch	256x256	0.8473	0.0574	weights
	DeepLabV3_ResNet101	Pretrained on COCO. Backbone frezed except layer_4	512x512	0.8935	0.0387	weights

Note

These models were trained with A100 40GB, A100 80GB GPUs and data shuffle, so if you train them again you may obtain different results also depending on your hardware. Of course there may be a better hyper-parameters configuration, this is not necessarily the best.