ReDO: Unsupervised Object Segmentation by Redrawing
We discover meaningful segmentation masks by redrawing regions of the images independently.
Table of Contents
- Random samples
- Samples for Flowers, LFW, CUB and toy dataset
- A more diverse dataset with two classes
- Datasets instructions
- Pretrained models
- Training ReDO
Samples for Flowers, LFW, CUB and toy dataset
A more diverse dataset with two classes
During the rebuttal process, we were asked to demonstrate that ReDO can work when the dataset contains multiple classes. We build a new dataset by combining LFW and Flowers images (without labels). This new dataset has more variability, contains different types of objects, and display a more obvious correlation between the object and the background. We trained ReDO without further hyperparameter tuning (not optimal), and obtained a reasonable accuracy of 0.856 and IoU of 0.691.
- Download and extract: Dataset, Segmentations, and data splits from http://www.robots.ox.ac.uk/~vgg/data/flowers/102/
- The obtained jpg folder, segmin folder and setid.mat file should be placed in the same data root folder.
- Download and extract Images and Segmentations from http://www.vision.caltech.edu/visipedia/CUB-200-2011.html
- Place the segmentations folder in the CUB_200_2011/CUB_200_2011 subfolder.
- Place the train_val_test_split.txt file from this repo in the CUB_200_2011/CUB_200_2011 subfolder.
- dataroot should be set to the CUB_200_2011/CUB_200_2011 subfolder.
- Download and extract the funneled images from http://vis-www.cs.umass.edu/lfw/
- Download and extract the ground truth images from http://vis-www.cs.umass.edu/lfw/part_labels/
- Place the obtained lfw_funneled and parts_lfw_funneled_gt_images folders in the same data root folder.
- Also place the train.txt, val.txt and test.txt files from the repo in this data root folder.
Tested on python3.7 with pytorch 1.0.1
Load pretrained models
Weights pretrained on Flowers, LFW, and CUB datasets can be downloaded from google drive.
- dataset_nets_state.tar.gz: pretrained weights for all 4 networks used during training in a single file.
The weights for the individual networks are also available, for instance if you only need to segment and/or redraw:
- dataset_netM_state.tar.gz: pretrained weights for mask extractor only. Enough if interested only in segmentation.
- dataset_netX_state.tar.gz: pretrained weights for region generators. Used to redraw objects.
- dataset_netD_state.tar.gz: pretrained weights for discriminator.
- dataset_netZ_state.tar.gz: pretrained weights for the network that infer the latent code z from image.
.tar.gz archives have to be uncompressed first to recover the .pth files containing the weights.
Provided example script needs at least netM and netX and is used as follows:
If using dataset_nets_state.pth on GPU cuda device 0
python example_load_pretrained.py --statePath path_to_nets_state.pth --dataroot path_to_data --device cuda:0
If using dataset_netX_state.pth and dataset_netM_state.pth on cpu:
python example_load_pretrained.py --statePathX path_to_netX_state.pth --statePathM path_to_netM_state.pth --dataroot path_to_data --device cpu
Training from scratch
python train.py --dataset flowers --nfX 32 --useSelfAttG --useSelfAttD --outf path_to_output_folder --dataroot path_to_data_folder --clean
python train.py --dataset lfw --useSelfAttG --useSelfAttD --outf path_to_output_folder --dataroot path_to_data_folder --clean
Some clarifications about the training process and the collapse issue:
As mentionned in the paper, the model can collapse with one region taking the whole image. This happens early in the training (at about 3-5k iterations) in some runs (about 3.5 out of 10 in my experiments). In this case, it is possible to restart training automatically using option --autoRestart .1 (for instance).
After these early stages, training should be stable. I stop training in the 20k~40k range, but the model gets unstable again if you train for too long.