Concrete results on COCO

PseudoSeg: Designing Pseudo Labels for Semantic Segmentation

PseudoSeg is a simple consistency training framework for semi-supervised image semantic segmentation, which has a simple and novel re-design of pseudo-labeling to generate well-calibrated structured pseudo labels for training with unlabeled or weakly-labeled data. It is implemented by Yuliang Zou (research intern) in 2020 Summer.

This is not an official Google product.

Instruction

Installation

Use a virtual environment

virtualenv -p python3 --system-site-packages env
source env/bin/activate

Install packages

pip install -r requirements.txt

Dataset

Create a dataset folder under the ROOT directory, then download the pre-created tfrecords for voc12 and coco, and extract them in dataset folder. You may also want to check the filenames for each split under data_splits folder.

Training

NOTE:

We train all our models using 16 V100 GPUs.
The ImageNet pre-trained models can be download here.
For VOC12, ${SPLIT} can be 2_clean, 4_clean, 8_clean, 16_clean_3 (representing 1/2, 1/4, 1/8, and 1/16 splits), NUM_ITERATIONS should be set to 30000.
For COCO, ${SPLIT} can be 32_all, 64_all, 128_all, 256_all, 512_all (representing 1/32, 1/64, 1/128, 1/256, and 1/512 splits), NUM_ITERATIONS should be set to 200000.

Supervised baseline

python train_sup.py \
  --logtostderr \
  --train_split="${SPLIT}" \
  --model_variant="xception_65" \
  --atrous_rates=6 \
  --atrous_rates=12 \
  --atrous_rates=18 \
  --output_stride=16 \
  --decoder_output_stride=4 \
  --train_crop_size="513,513" \
  --num_clones=16 \
  --train_batch_size=64 \
  --training_number_of_steps="${NUM_ITERATIONS}" \
  --fine_tune_batch_norm=true \
  --tf_initial_checkpoint="${INIT_FOLDER}/xception_65/model.ckpt" \
  --train_logdir="${TRAIN_LOGDIR}" \
  --dataset_dir="${DATASET}"

PseudoSeg (w/ unlabeled data)

python train_wss.py \
  --logtostderr \
  --train_split="${SPLIT}" \
  --train_split_cls="train_aug" \
  --model_variant="xception_65" \
  --atrous_rates=6 \
  --atrous_rates=12 \
  --atrous_rates=18 \
  --output_stride=16 \
  --decoder_output_stride=4 \
  --train_crop_size="513,513" \
  --num_clones=16 \
  --train_batch_size=64 \
  --training_number_of_steps="${NUM_ITERATIONS}" \
  --fine_tune_batch_norm=true \
  --tf_initial_checkpoint="${INIT_FOLDER}/xception_65/model.ckpt" \
  --train_logdir="${TRAIN_LOGDIR}" \
  --dataset_dir="${DATASET}"

PseudoSeg (w/ image-level labeled data)

python train_wss.py \
  --logtostderr \
  --train_split="${SPLIT}" \
  --train_split_cls="train_aug" \
  --model_variant="xception_65" \
  --atrous_rates=6 \
  --atrous_rates=12 \
  --atrous_rates=18 \
  --output_stride=16 \
  --decoder_output_stride=4 \
  --train_crop_size="513,513" \
  --num_clones=16 \
  --train_batch_size=64 \
  --training_number_of_steps="${NUM_ITERATIONS}" \
  --fine_tune_batch_norm=true \
  --tf_initial_checkpoint="${INIT_FOLDER}/xception_65/model.ckpt" \
  --train_logdir="${TRAIN_LOGDIR}" \
  --dataset_dir="${DATASET}" \
  --weakly=true

Evaluation

NOTE: ${EVAL_CROP_SIZE} should be 513,513 for VOC12, 641,641 for COCO.

python eval.py \
  --logtostderr \
  --eval_split="val" \
  --model_variant="xception_65" \
  --atrous_rates=6 \
  --atrous_rates=12 \
  --atrous_rates=18 \
  --output_stride=16 \
  --decoder_output_stride=4 \
  --eval_crop_size="${EVAL_CROP_SIZE}" \
  --checkpoint_dir="${TRAIN_LOGDIR}" \
  --eval_logdir="${EVAL_LOGDIR}" \
  --dataset_dir="${DATASET}" \
  --max_number_of_evaluations=1

Visualization

NOTE: ${VIS_CROP_SIZE} should be 513,513 for VOC12, 641,641 for COCO.

python vis.py \
  --logtostderr \
  --vis_split="val" \
  --model_variant="xception_65" \
  --atrous_rates=6 \
  --atrous_rates=12 \
  --atrous_rates=18 \
  --output_stride=16 \
  --decoder_output_stride=4 \
  --vis_crop_size="${VIS_CROP_SIZE}" \
  --checkpoint_dir="${CKPT}" \
  --vis_logdir="${VIS_LOGDIR}" \
  --dataset_dir="${PASCAL_DATASET}" \
  --also_save_raw_predictions=true

Citation

If you use this work for your research, please cite our paper.

@article{zou2020pseudoseg,
  title={PseudoSeg: Designing Pseudo Labels for Semantic Segmentation},
  author={Zou, Yuliang and Zhang, Zizhao and Zhang, Han and Li, Chun-Liang and Bian, Xiao and Huang, Jia-Bin and Pfister, Tomas},
  journal={International Conference on Learning Representations (ICLR)},
  year={2021}
}

	_PASCAL_VOC_SEG_INFORMATION = DatasetDescriptor(
	splits_to_sizes={
	'train': 1464,
	'train_aug': 10582,
	'trainval': 2913,
	'val': 1449,
	# Splits for semi-supervised
	'4_clean': 366,
	'8_clean': 183,
	# Balanced 1/16 split
	# Sample with rejection, suffix represents the sample index
	# e.g., 16_clean_3 represents the 3rd random shuffle to sample 1/16
	# split, given a fixed random seed 8888
	'16_clean_3': 92,
	'16_clean_14': 92,
	'16_clean_20': 92,
	# More images
	'coco2voc_labeled': 91937,
	'coco2voc_unlabeled': 215340,
	},

	brightness = 0.5
	contrast = 0.5
	saturation = 0.5
	hue = 0.25

googleinterns / wss Goto Github PK

wss's Introduction

Instruction

Installation

Dataset

Training

Supervised baseline

PseudoSeg (w/ unlabeled data)

PseudoSeg (w/ image-level labeled data)

Evaluation

Visualization

Citation

wss's People

Contributors

Stargazers

Watchers

Forkers

wss's Issues

Recommend Projects

Recommend Topics

Recommend Org