This code is pytorch implementation of Semantic Instance Segmentation with a Discriminative Loss Function (https://arxiv.org/abs/1708.02551) with CVPPP-2017 Dataset (https://www.plant-phenotyping.org/CVPPP2017)

I slightly updated and revised the code of https://github.com/Wizaron/instance-segmentation-pytorch which does not match with python3. The updated content is as follows.

• compatibility with python3

• revised the dataset which does not use lmdb

• added other networks (UNet, UNet with CBAM, DeepLabV3)

  • unfortunately semantic & instance segmentation result of DeepLabV3 are very bad

• ReSeg with CoordConv(with r) is possible

My paper 'Leaf Instance Segmentation with Attention Based U-Net & Discriminative Loss' which utilizes this code was submitted for participation at 'Summer Annual conference of IEIE, 2022 (https://conf.theieie.org/2022s/)'

• You should execute python -m visdom.server before training

• train3.py -> pred_list2.py -> evaluation.ipynb

• I couldn't solve the error of training Stacked Recurrent Hourglass. The training does not proceed from 2 epochs due to errors in the back-propagation process.

• Updated(23/06/19)

  • If you use the latest version of pytorch (torch==2.1.1+cu121), please use train3_update.py & model_update.py instead of train3.py & model.py
    • .next() method was deprecated (line 257 of model.py)

• Instance Counter : predicts normalized # of leaf instances
• Semantic Head : predicts semantic mask (f.g / b.g)
• Instance Head : predicts 32 dims embedding space which has to be clustered by K-Means

input / pred / GT

ReNet (used in ReSeg) : https://arxiv.org/abs/1505.00393

ReSeg : https://arxiv.org/abs/1511.07053

SegNet : https://arxiv.org/pdf/1511.00561.pdf

CBAM (Convolutional Block Attention Module) : https://arxiv.org/abs/1807.06521