Effective Guidance for Model Attention with Simple Yes-no Annotations

Installation

You can download all required packages by using the following command.

conda env create -f environment.yml

Dataset

You can download each dataset by clicking the dataset name.

• Waterbirds
• Biased CelebA
• Backgrounds Challenge

Pretrained Models

You can download the pretrained models that attend to irrelevant areas at this link. Please download and place the models under the folder model.

Saliency Maps for Pretrained Models

You can download the files for the saliency maps for the pretrained models at this link. Please download and place them under the folder interpretation.

File structure

We suggest you to save data, models, and interpretation results following the file structure below.

.
├─ data
│   ├─ waterbirds
│   │   ├─ images: contains 200 folders; each folder for each bird species
│   │   └─ metadata.csv
│   ├─ biased_celeba
│   │   ├─ celeba
│   │   │   └─ img_align_celeba: this folder contains images
│   │   ├─ list_attr_celeba.txt
│   │   ├─ biased_celeba_black_hair_1_100_train.csv
│   │   └─ biased_celeba_black_hair_test.csv
│   └─ background
│   │   ├─ original
│   │   │   ├─ train: contains 9 folders; each folder for each class
│   │   │   └─ val: contains 9 folders; each folder for each class
│   │   ├─ mixed_same
│   │   │   └─ val: contains 9 folders; each folder for each class
│   │   ├─ mixed_rand
│   │   │   └─ val: contains 9 folders; each folder for each class
│   │   └─ only_fg
│   │   │   └─ val: contains 9 folders; each folder for each class
├─ model
│   ├─ waterbirds_resnet50.pth
│   ├─ biased_celeba_resnet50.pth
│   └─ background_resnet50.pt
├─ interpretation
│   ├─ waterbirds_gradcam.json
│   ├─ biased_celeba_gradcam.json
│   └─ background_gradcam.json
├─ annotation
├─ performance
├─ script
├─ src
└─ README.md

Running CRAYON

On the command line, type as following:

cd script/DATA_NAME
sh METHOD_NAME.sh

For example, if you want to run CRAYON-Attention for the Waterbirds dataset, you can run the code below:

cd script/waterbirds
sh crayon_att.sh

Evaluation

CRAYON achieves state-of-the-art performance, outperforming 12 methods across 3 benchmark datasets, surpassing approaches that require more complex annotations. You can see more details about the training configurations of CRAYON and compared methods at this link.

Waterbirds

Method	Annotation	WGA	MGA
Original (Unrefined)	-	28.35	72.08
CRAYON-Attention (ours)	Yes-No	72.31	85.23
CRAYON-Pruning (ours)	Yes-No	68.97	83.13
CRAYON-Pruning+Attention (ours)	Yes-No	76.04	86.03
JtT	-	46.88	78.29
MaskTune	-	45.67	79.13
LfF	-	44.64	77.24
SoftCon	-	46.10	79.93
FLAC	-	40.68	80.77
LC	-	61.65	80.43
CnC	-	46.98	77.80
RRR	Map	53.96	82.29
GradMask	Map	58.38	82.78
ActDiff	Map	64.58	84.54
GradIA	Yes-No, Map	60.87	83.17
Bounding Box	Bounding Box	66.36	85.85

Biased CelebA

Method	Annotation	WGA	MGA
Original (Unrefined)	-	32.60	73.71
CRAYON-Attention (ours)	Yes-No	83.29	89.61
CRAYON-Pruning (ours)	Yes-No	69.51	86.75
CRAYON-Pruning+Attention (ours)	Yes-No	84.38	90.13
JtT	-	35.25	74.29
MaskTune	-	37.72	78.85
LfF	-	44.35	77.05
SoftCon	-	42.38	76.16
FLAC	-	39.31	76.73
LC	-	64.98	84.80
CnC	-	37.76	75.34
RRR	Map	42.16	78.64
GradMask	Map	8.43	65.72
ActDiff	Map	41.29	79.08
GradIA	Yes-No, Map	41.80	76.58
Bounding Box	Bounding Box	33.34	73.03

Backgrounds Challenge

We note that lower BG-Gap means better performance.

Method	Annotation	MR	BG-Gap
Original (Unrefined)	-	78.27	12.99
CRAYON-Attention (ours)	Yes-No	80.85	8.52
CRAYON-Pruning (ours)	Yes-No	78.61	12.18
CRAYON-Pruning+Attention (ours)	Yes-No	81.66	8.40
JtT	-	77.61	12.99
MaskTune	-	78.14	12.54
LfF	-	78.23	12.41
SoftCon	-	73.14	11.58
FLAC	-	79.91	9.75
LC	-	74.78	13.35
CnC	-	77.87	12.91
RRR	Map	80.08	10.68
GradMask	Map	80.11	10.63
ActDiff	Map	75.84	11.73
GradIA	Yes-No, Map	79.79	11.26
Bounding Box	Bounding Box	80.82	9.81

Varying Number of Annotations

We evaluate how the number of annotations affects CRAYON's performance.

Waterbirds
Biased CelebA
Backgrounds Challenge

Hyperparameter Sensitivity

We examine how the two hyperparameters, $\alpha$ and $\beta$, affect the performance of CRAYON-Attention. While it shows effectiveness across various values of $\alpha$ and $\beta$, careful tuning can achieve much higher performance.

Waterbirds
Biased CelebA
Backgrounds Challenge

Likewise, we also examine the effect on CRAYON-Pruning+Attention.

Waterbirds
Biased CelebA
Backgrounds Challenge