VitroVision U-Net Small (greenhouse v2)

Segments a tissue-culture plant from a bottle image. U-Net + MobileNetV3-Small encoder.

Training

• Arch: smp.Unet(encoder_name="timm-mobilenetv3_small_100", in_channels=3, classes=1) (~3.6M params)
• Dataset: Project-AgML/greenhouse_leafy_segmentation (1200 pairs -> train 1080 / val 120)
• Loss: BCE + Dice
• Val Dice: 0.9817
• Ready-vote threshold READY_HEIGHT=0.20 (Youden-balanced, tuned on 98 labeled bottle images)

Limitations (honest)

• Trained on a public greenhouse dataset; not fine-tuned on the 100-bottle in-vitro test set (that set is test/eval only).
• 2D projected traits (coverage / height / width proxy) discriminate the readiness verdict only modestly (AUC ~0.64) -> motivates 3D / multi-trait work.
• Prototype, not a production grader.

Load

import torch, segmentation_models_pytorch as smp
m = smp.Unet(encoder_name="timm-mobilenetv3_small_100", in_channels=3, classes=1)
m.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"))