Single-Stage Semantic Segmentation from Image Labels

This repository contains the original implementation of our paper:

Single-stage Semantic Segmentation from Image Labels
Nikita Araslanov and Stefan Roth
CVPR 2020. [pdf] [supp] [arXiv]

Contact: Nikita Araslanov [email protected]

We attain competitive results by training a single network model for segmentation in a self-supervised fashion using only image-level annotations (one run of 20 epochs on Pascal VOC).

Setup

0. Minimum requirements. This project was originally developed with Python 3.6, PyTorch 1.0 and CUDA 9.0. The training requires at least two Titan X GPUs (12Gb memory each).

1. Setup your Python environment. Please, clone the repository and install the dependencies. We recommend using Anaconda 3 distribution:

   conda create -n <environment_name> --file requirements.txt
   

2. Download and link to the dataset. We train our model on the original Pascal VOC 2012 augmented with the SBD data (10K images in total). Download the data from:

   • VOC: Training/Validation (2GB .tar file)
   • SBD: Training (1.4GB .tgz file)

   Link to the data:

   ln -s <your_path_to_voc> <project>/data/voc
   ln -s <your_path_to_sbd> <project>/data/sbd
   

   Make sure that the first directory in data/voc is VOCdevkit; the first directory in data/sbd is benchmark_RELEASE.

3. Download pre-trained models. Download the initial weights (pre-trained on ImageNet) for the backbones you are planning to use and place them into <project>/models/weights/.

   Backbone	Initial Weights	Comment
   WideResNet38	ilsvrc-cls_rna-a1_cls1000_ep-0001.pth (402M)	Converted from mxnet
   VGG16	vgg16_20M.pth (79M)	Converted from Caffe
   ResNet50	resnet50-19c8e357.pth	PyTorch official
   ResNet101	resnet101-5d3b4d8f.pth	PyTorch official

Training, Inference and Evaluation

The directory launch contains template bash scripts for training, inference and evaluation.

Training. For each run, you need to specify names of two variables, for example

EXP=baselines
RUN_ID=v01

Running bash ./launch/run_voc_resnet38.sh will create a directory ./logs/pascal_voc/baselines/v01 with tensorboard events and will save snapshots into ./snapshots/pascal_voc/baselines/v01.

Inference. To generate final masks, please, use the script ./launch/infer_val.sh. You will need to specify:

• EXP and RUN_ID you used for training;
• OUTPUT_DIR the path where to save the masks;
• FILELIST specifies the file to the data split;
• SNAPSHOT specifies the model suffix in the format e000Xs0.000. For example, e020Xs0.928;
• (optionally) EXTRA_ARGS specify additional arguments to the inference script.

Evaluation. To compute IoU of the masks, please, run ./launch/eval_seg.sh. You will need to specify SAVE_DIR that contains the masks and FILELIST specifying the split for evaluation.

Pre-trained model

For testing, we provide our pre-trained WideResNet38 model:

Backbone	Val	Val (+ CRF)	Link
WideResNet38	59.7	62.7	model_enc_e020Xs0.928.pth (527M)

The also release the masks predicted by this model:

Split	IoU	IoU (+ CRF)	Link	Comment
train-clean (VOC+SBD)	64.7	66.9	train_results_clean.tgz (2.9G)	Reported IoU is for VOC
val-clean	63.4	65.3	val_results_clean.tgz (423M)
val	59.7	62.7	val_results.tgz (427M)
test	62.7	64.3	test_results.tgz (368M)

The suffix -clean means we used ground-truth image-level labels to remove masks of the categories not present in the image. These masks are commonly used as pseudo ground truth to train another segmentation model in fully supervised regime.

Acknowledgements

We thank PyTorch team, and Jiwoon Ahn for releasing his code that helped in the early stages of this project.

Citation

We hope that you find this work useful. If you would like to acknowledge us, please, use the following citation:

@InProceedings{Araslanov:2020:SSS,
author = {Araslanov, Nikita and Roth, Stefan},
title = {Single-Stage Semantic Segmentation From Image Labels},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
pages = {4253--4262}
year = {2020}
}