jlchen-c/semformer

SemFormer

The official code for SemFormer: Semantic Guided Activation Transformer for Weakly Supervised Semantic Segmentation.

Runtime Environment

• Python 3.6
• PyTorch 1.7.1
• CUDA 11.0
• 2 x NVIDIA A100 GPUs
• more in requirements.txt

Usage

Install python dependencies

python -m pip install -r requirements.txt

Download PASCAL VOC 2012 devkit

Follow instructions in http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#devkit.

Train and evaluate the model.

1. Train SemFormer for generating CAMs

1.1 Train CAAE.

CUDA_VISIBLE_DEVICES=0,1 python train_caae.py --tag CAAE@DeiT-B-Dist

1.2 Train SemFormer.

CUDA_VISIBLE_DEVICES=0,1 python train_semformer.py --tag SemFormer@CAAE@DeiT-B-Dist

Or use the checkpoint we porvide in experiments/models/SemFormer@[email protected].

2. Inference SemFormer for generating CAMs

CUDA_VISIBLE_DEVICES=0 python inference_semformer.py --tag SemFormer@CAAE@DeiT-B-Dist --domain train_aug

Evaluate CAMs. [optinal]

python evaluate.py --experiment_name SemFormer@CAAE@DeiT-B-Dist@train@scale=0.5,1.0,1.5,2.0 --domain train

3. Apply Random Walk (RW) to refine the generated CAMs

2.1. Make affinity labels to train AffinityNet.

python make_affinity_labels.py --experiment_name SemFormer@CAAE@DeiT-B-Dist@train@scale=0.5,1.0,1.5,2.0 --domain train_aug

2.2. Train AffinityNet using the generated affinity labels.

CUDA_VISIBLE_DEVICES=0,1 python train_affinitynet.py --tag AffinityNet@SemFormer --label_name SemFormer@CAAE@DeiT-B-Dist@train@scale=0.5,1.0,1.5,2.0@aff_fg=0.11_bg=0.15

4. Make pseudo labels.

4.1 Inference random walk (affinitynet) to refine the generated CAMs.

CUDA_VISIBLE_DEVICES=0 python inference_rw.py --model_name AffinityNet@SemFormer --cam_dir SemFormer@CAAE@DeiT-B-Dist@train@scale=0.5,1.0,1.5,2.0 --domain train_aug

4.2 Apply CRF to generate pseudo labels.

python make_pseudo_labels.py --experiment_name AffinityNet@SemFormer@train@beta=10@exp_times=8@rw --domain train_aug --crf_iteration 1

5. Train and Evaluate the segmentation model using the pseudo labels

Please follow the instructions in this repo to train and evaluate the segmentation model.

6. Results

Qualitative segmentation results on PASCAL VOC 2012 (mIoU (%)). Supervision: pixel-level ($\mathcal{F}$), box-level ($\mathcal{B}$), saliency-level ($\mathcal{S}$), and image-level ($\mathcal{I}$).

Method	Publication	Supervision	val	test
DeepLabV1	ICLR'15	$\mathcal{F}$	68.7	71.6
DeepLabV2	TPAMI'18	$\mathcal{F}$	77.7	79.7
BCM	CVPR'19	$\mathcal{I} + \mathcal{B}$	70.2	-
BBAM	CVPR'21	$\mathcal{I} + \mathcal{B}$	73.7	73.7
ICD	CVPR'20	$\mathcal{I} + \mathcal{S}$	67.8	68.0
EPS	CVPR'21	$\mathcal{I} + \mathcal{S}$	71.0	71.8
BES	ECCV'20	$\mathcal{I}$	65.7	66.6
CONTA	NeurIPS'20	$\mathcal{I}$	66.1	66.7
AdvCAM	CVPR'21	$\mathcal{I}$	68.1	68.0
OC-CSE	ICCV'21	$\mathcal{I}$	68.4	68.2
RIB	NeurIPS'21	$\mathcal{I}$	68.3	68.6
CLIMS	CVPR'22	$\mathcal{I}$	70.4	70.0
MCTFormer	CVPR'22	$\mathcal{I}$	71.9	71.6
SemFormer (ours)	-	$\mathcal{I}$	73.7	73.2

Acknowledgement

This repo is modified from Puzzle-CAM, thanks for their contribution to the community.

semformer's Issues

Segmentation Model

Hi, I'm amazing about the results of your paper, so I've focused on your work and tried to reproduce it. I'd like to ask you some questions about the segmentation network. In your repo, you only mention the segmentation model's training model but do not provide the specific training parameters or the method of post-processing. When I reproduce it again, I can't reproduce the results you gave in your paper. Could you please provide the specific parameters of segmentation retrain? 

FileNotFoundError: [Errno 2] No such file or directory: './VOC2012/Annotations/2011_003194.xml'

您好，我在运行代码训练train_caae.py时发现VOC2012数据集中annotations中有些xml文件是没有的，想问一下您这边训练这个问题是怎么处理的呢？

Questions for Mask mIoU

Greetings!

Recently I have read your work SemFormer, and deeply surprised by the amazing results! So I find this repo and try to reproduce the results.

I meet a problem that I could not get the 73.2% mIoU mask reported in your paper by a 63.7% mIoU CAM. Could you share your hyper parameters of PSA and other post process? Thanks a lot!