seg: tools for evaluating binary semantic image segmentation models

• Frameworks & Models:
  • torchvision
  • segmentation_models.pytorch
  • Mask2Former
  • segment-anything (SAM)
  • SAM2
  • MobileSAM
  • UNet (segmentation_of_roots_in_soil_with_unet)
  • RootNav 2.0
  • segroot
• Automatic mixed precision training
• On-the-fly loading images from disk for training
• Support for arbitrary patch sizes and some augmentation

Citation

If you use this work, please cite:

A systematic comparison of transformers and ConvNets for root segmentation across nine datasets.

Smith, Abraham George, et al. "A systematic comparison of transformers and ConvNets for root segmentation across nine datasets." Plant Methods (2026).

@article{smith2026systematic,
  title={A systematic comparison of transformers and ConvNets for root segmentation across nine datasets},
  author={Smith, Abraham George and Lamprinidis, Sotiris and Seethepalli, Anand and York, Larry M and Han, Eusun and M{\"o}hl, Patrick and Boulata, Kyriaki and Thorup-Kristensen, Kristian and Petersen, Jens},
  journal={Plant Methods},
  year={2026},
  publisher={Springer}
}

Installation

Some of the included packages have different requirements. E.g. in an Ubuntu box, one will need g++ and python3-dev.

IMPORTANT Some Mask2Former custom ops require torch == 2.5.1

We need some custom steps wrt detectron and mask2former ./install.sh

Download the pretrained models ./download_pretrained.sh

Download the datasets cd data; for f in *_import.py; do ./$f; done

Usage

The data folder should contain a "train" and "val" directory, each containing an "annotations" and "images":

- data
  data/train
  data/train/annotations
  data/train/annotations/0000.png
  ...
  data/train
  data/train/images
  data/train/images/0000.png
  ...
  data/val
  data/val/annotations
  data/val/annotations/0000.png
  ...
  data/val
  data/val/images
  data/val/images/0000.png
  ...

The samples will be cropped to -s,--size if larger (default: 1024 x 1024):

Run an experiment using: ./seg.py $data_path unet/vanilla -o output

Then in the output directory we get:

• results: A text file with the epoch, step, seconds per epoch, and best f1 score achieved
• *-img.png: A sample image from the validation set
• *-pred.png: The prediction of the model for the sample image
• *-true.png: The ground truth for the segmentation
• checkpoint_best.pth: The best model
• config: The training hyper-parameters

Command-line options

options:
  -h, --help            show this help message and exit
  -a LEARNING_RATE, --learning-rate LEARNING_RATE
  -b BATCH_SIZE, --batch-size BATCH_SIZE
  -e EPOCHS, --epochs EPOCHS
  -f EVAL_FREQUENCY, --eval-frequency EVAL_FREQUENCY
  -j NUM_WORKERS, --num-workers NUM_WORKERS
  -m, --mixed-precision
  -p PATIENCE, --patience PATIENCE
  -o CHECKPOINT_DIR, --checkpoint-dir CHECKPOINT_DIR
  -s SHAPE SHAPE, --shape SHAPE SHAPE
  -t TIMEOUT, --timeout TIMEOUT
  -w WARMUP_STEPS_MUL, --warmup-steps-mul WARMUP_STEPS_MUL
  -A, --sparse-annotations
  -C, --clip-gradients
  -N, --no-optimizations
  -P, --pretrained
  -S, --save-val-images
  -X, --extra-val-metrics
  -w WARMUP_STEPS, --warmup-steps WARMUP_STEPS

Developing

Every model file should define

• A models dict with keys the name of each model
• A new(name, pretrained: Bool) function that returns a model
• A model that takes a (N x 3 x H x W) image and returns a (N x 1 x H x W) prediction

Models

m2f/R101
m2f/R50
m2f/swin-base
m2f/swin-small
m2f/swin-tiny
mb_sam/vit_t
sam/vit-base
sam/vit-huge
sam/vit-large
segmentation_pytorch/deeplabv3+-ENCODER
segmentation_pytorch/deeplabv3-ENCODER
segmentation_pytorch/linknet-ENCODER
segmentation_pytorch/manet-ENCODER
segmentation_pytorch/pan-ENCODER
segmentation_pytorch/pspnet-ENCODER
segmentation_pytorch/segformer-ENCODER
segmentation_pytorch/unet++-ENCODER
segmentation_pytorch/unet-ENCODER

ENCODER:
        densenet{121,161,169,201}
        efficientnet-b{0..6}
        inceptionv4
        mobilenet_v2
        mobileone_s{0..4}
        resnet{18,50,101}
        timm-efficientnet-b{0..6}
        vgg{11,19}
        xception

torchvision_models/deeplabv3-mobilenet_v3_large
torchvision_models/deeplabv3-resnet101
torchvision_models/deeplabv3-resnet50
torchvision_models/fcn-resnet101
torchvision_models/fcn-resnet50
torchvision_models/lraspp-mobilenet_v3_large
unet/root_painter
unet/vanilla