Home

Awesome

ACDRNet

Official PyTorch implementation of "End to End Trainable Active Contours via Differentiable Rendering" (link)

Prerequisites

Data preparation

Buildings

The data folder should have the following format:

images/
    AAA.xxx
    BBB.xxx
masks/
    AAA.xxx
    BBB.xxx

Download links:

Cityscapes

Convert Cityscapes images to HDF5 for fast crop inference:

python data/make_hdf5.py [--images-path] [--outdir]

Generate Cityscapes instances .json files:

python data/generate_cityscapes_instances.py [--outdir] [--outfile] [--min-area]

Make sure you use the same path in both files for --outdir. The data folder should have the following format:

cityscapes_instances/
    train/
        all_classes_instances.json
    train_val/
        all_classes_instances.json
    val/
        all_classes_instances.json
    all_images.hdf5

Download links:

Train and Evaluate

Use the train.py for training. You can set the "validation" dataset to the "test" dataset for easy evaluation every --eval-rate epochs.

Training options:

# Training
--epochs        Number of epochs
--start-epoch   Starting epoch
--batch-size    Batch size
--lr            Learning rate
--lr-step       LR scheduler step

# Architecture
--arch          Network architecture. "unet" or "resnet"
--image-size    Neural Renderer output size
--dec-size      Spatial size of the decoder. Only relevant for ResNet
--enc-dim       Encoder dim(channels). Only relevant for UNet
--dec-dim       Decoder dim(channels)
--stages        ResNet skip connections
--drop          Dropout rate

# Active contour
--num-nodes     Number of nodes
--iter          AC number of iterations
--lmd-balloon   Balloon
--lmd-curve     Curvature
--lmd-dist      Distance

# Data
--train-dataset Training dataset
--ann-train     Split for training
--ann-val       Split tor evaluation

# Cityscapes Data
--inst-path     Path to Cityscapes instances directory
--ann-type      Type of annotation, full instance or only components. 
--class-name    Class for Cityscapes dataset
--loops         Data repetition in Cityscapes dataset

# Buildings Data
--data-path     Path to buildings dataset directory

# Misc
--eval-rate     Evaluate after "eval_rate" epochs
--save-rate     Save rate is "save_rate" * "eval_rate"
--checkname     Checkname
--resume        Resume file path

Cityscapes example:

python train.py --arch resnet --class-name train [--inst-path] --ann-type full --train-dataset cityscapes

Tunning