pp-seg

Potted Plants Segmentation - binary plant segmentation in a greenhouse setting.

Dataset

Public dataset from Roboflow, collected in a greenhouse and annotated in COCO segmentation format (greenhouse-3-1/). After removing unannotated images: 695 train / 208 valid / 106 test. All images resized to 512x512. Foreground covers only ~2.5% of pixels, which drives most of the design choices (loss, metrics).

Approach

1. Baseline - smp.Unet (ResNet34, ImageNet pretrained). Strong but 24.4M params.
2. BinarySegCNN - flat encoder-decoder, 185K params. Struggles with thin structures (stems, partial leaves); changing the loss doesn't fix it, so the bottleneck is the architecture.
3. BinaryUNet - custom UNet with skip connections, configurable depth and width. Trained with BCE + Dice loss to handle the class imbalance.
4. Grid search - 3-fold CV over lr, batch size, base filters, depth, selecting on plant IoU. Winner: base_filters=32, depth=2, lr=1e-3, batch_size=8.

Results

Test-set plant IoU, threshold tuned on the validation split:

Model	Params	Plant IoU
smp.Unet (ResNet34, ImageNet)	24.4M	0.9102
BinarySegCNN (BCEDice)	185K	0.8418
BinaryUNet f16 d3 (BCEDice)	488K	0.9079
BinaryUNet f32 d2 (final)	472K	0.9092

Structure

• ppseg.ipynb - data exploration, preprocessing, models, grid search, final training and evaluation, results summary
• models/ - BinarySegCNN, BinaryUNet
• utils/ - dataset and loaders (data.py), training loop (train.py), losses (losses.py), evaluation protocol (evaluate.py), plotting (viz.py), grid search (kfold_gridsearch.py)
• outputs/ - checkpoints, normalization stats, grid search results (gitignored)

Running it

Sections of the notebook are independent: each one creates its own dataloaders, so you can run any section from a fresh kernel without running the ones before it. Training cells are the slow part; every trained model is evaluated from its checkpoint, so evaluation cells can be re-run without retraining.

Future work

• Improve generalization with a bigger dataset covering different plant species
• Revisit the pretrained route, which should generalize better with more varied data
• Build a new dataset and fine-tune the model on pseudo-labels generated with SAM 3