zhouhuan-hust / Open-World-Semantic-Segmentation

Code for ICCV2021 paper "Deep Metric Learning for Open World Semantic Segmentation".

DMLNet for open world semantic segmentation

This code is for ICCV2021 paper: "Deep Metric Learning for OpenWorld Semantic Segmentation".

The open-set semantic segmentation module is in ./anomaly and the incremental few-shot learning module is in ./DeepLabV3Plus-Pytorch.

Open-set semantic segmentation module

We provide the procedure to reproduce the DML results of Table 1 in the paper.

Dataset

Download official StreetHazards training dataset and test dataset, then extract it as following:

YourPath/Open_world
            /DeepLabV3Plus-Pytorch
            /anomaly
            /data
                /streethazards
                    /train
                    /test

Pretrained model

Download pretrained model checkpoints, and put them under ./anomaly/ckpt/.

Evaluation

All codes of open-set semantic segmentation are in ./anomaly, so first:

cd anomaly

Then set the dataset route:

vim config/test_ood_street.yaml
change the root_dataset, list_train and list_val as:
  root_dataset: "YourPath/Open_world/data/streethazards/test"
  list_train: "YourPath/Open_world/data/streethazards/train/train.odgt"
  list_val: "YourPath/Open_world/data/streethazards/test/test.odgt"

Now evaluate:

python eval_ood_traditional.py --ood dissum

Incremental few-shot learning module

We provide the procedure to reproduce '16+1 setting 5 shot' results of Table 4 in the paper.

Dataset

Download Cityscapes dataset and extract it as following:

YourPath/Open_world
            /DeepLabV3Plus-Pytorch
            /anomaly
            /data
                /streethazards
                    /train
                    /test
                /cityscapes
                    /gtFine
                    /leftImg8bit

Pretrained model

Download the pretrained models from checkpoints. Put 4 pth files into the ./DeepLabV3Plus-Pytorch/ckpt/, and prototype_car_5_shot.json into ./DeepLabV3Plus-Pytorch.

16+1 setting 5 shot evaluation

All codes of incremental few-shot learning module are in ./DeepLabV3Plus-Pytorch. So first:

cd DeepLabV3Plus-Pytorch

FT

python test_self_distillation.py --model deeplabv3plus_embedding_self_distillation_resnet101 --dataset cityscapes --gpu_id 0 --lr 0.1 --crop_size 768 --batch_size 5 --output_stride 16 --data_root ../data/cityscapes --total_itrs 10 --val_interval 10 --novel_cls 1 --ckpt ./ckpt/best_deeplabv3plus_embedding_self_distillation_resnet101_cityscapes_161_FT.pth --test_only

PLM_all

python test_self_distillation.py --model deeplabv3plus_embedding_self_distillation_resnet101 --dataset cityscapes --gpu_id 0 --lr 0.1 --crop_size 768 --batch_size 5 --output_stride 16 --data_root ../data/cityscapes --total_itrs 10 --val_interval 10 --novel_cls 1 --ckpt ./ckpt/best_deeplabv3plus_embedding_self_distillation_resnet101_cityscapes_161_5_shot_PLM.pth --test_only

PLM_latest

vim test_self_distillation.py
comment out line 292, 295-297
uncomment line 293

python test_self_distillation.py --model deeplabv3plus_embedding_self_distillation_resnet101 --dataset cityscapes --gpu_id 0 --lr 0.1 --crop_size 768 --batch_size 5 --output_stride 16 --data_root ../data/cityscapes --total_itrs 10 --val_interval 10 --novel_cls 1 --ckpt ./ckpt/best_deeplabv3plus_embedding_self_distillation_resnet101_cityscapes_161_5_shot_PLM.pth --test_only

NPM

python test_embedding.py --model deeplabv3plus_embedding_resnet101 --dataset cityscapes --gpu_id 0 --lr 0.1 --crop_size 768 --batch_size 16 --output_stride 16 --data_root ../data/cityscapes --test_only --ckpt ./ckpt/best_deeplabv3plus_embedding_resnet101_cityscapes_131415.pth

All 17

vim test_embedding.py
change line 661 opts.num_classes to 17
comment out line 428-451

python test_embedding.py --model deeplabv3plus_embedding_resnet101 --dataset cityscapes --gpu_id 0 --lr 0.1 --crop_size 768 --batch_size 16 --output_stride 16 --data_root ../data/cityscapes --test_only --ckpt ./ckpt/best_deeplabv3plus_embedding_resnet101_cityscapes_1415.pth

First 16

vim datasets/cityscapes.py
change line 71 unknown target to [13,14,15]
vim test_embedding.py
change line 661 opts.num_classes to 16
comment out line 428-451

python test_embedding.py --model deeplabv3plus_embedding_resnet101 --dataset cityscapes --gpu_id 0 --lr 0.1 --crop_size 768 --batch_size 16 --output_stride 16 --data_root ../data/cityscapes --test_only --ckpt ./ckpt/best_deeplabv3plus_embedding_resnet101_cityscapes_131415.pth