SCOUTER: Slot Attention-based Classifier for Explainable Image Recognition

Abstract

Explainable artificial intelligence has been gaining attention in the past few years. However, most existing methods are based on gradients or intermediate features, which are not directly involved in the decision-making process of the classifier. In this paper, we propose a slot attention-based classifier called SCOUTER for transparent yet accurate classification. Two major differences from other attention-based methods include: (a) SCOUTER's explanation is involved in the final confidence for each category, offering more intuitive interpretation, and (b) all the categories have their corresponding positive or negative explanation, which tells "why the image is of a certain category" or "why the image is not of a certain category." We design a new loss tailored for SCOUTER that controls the model's behavior to switch between positive and negative explanations, as well as the size of explanatory regions. Experimental results show that SCOUTER can give better visual explanations while keeping good accuracy on small and medium-sized datasets.

Model Structure

SCOUTER is built on top of the recently-emerged slot attention, which offers an object-centric approach for image representation. Based on this approach, we propose an explainable slot attention (xSlot) module. The output from the xSlot module is directly used as the confidence values for each category and thus commonly used fully-connected (FC) layer-based classifiers are no longer necessary. The whole network, including the backbone, is trained with the SCOUTER loss, which provides control over the size of explanatory regions and switching between positive and negative explanations.

Usage

Enable distributed training (if desired)

python -m torch.distributed.launch --nproc_per_node=4 --use_env train.py --world_size 4

Imagenet

Training for Imagenet dataset (Base Model)

python train.py --dataset ImageNet --model resnest26d --batch_size 70 --epochs 20 \
--num_classes 10 --use_slot false \
--vis false --channel 2048 --freeze_layers 0 \
--dataset_dir ../data/imagenet/ILSVRC/Data/CLS-LOC/

Positive Scouter for Imagenet dataset

python train.py --dataset ImageNet --model resnest26d --batch_size 70 --epochs 20 \
--num_classes 10 --use_slot true --use_pre false --loss_status 1 --slots_per_class 1 \
--power 2 --to_k_layer 3 --lambda_value 1 --vis false --channel 2048 --freeze_layers 0 \
--dataset_dir ../data/imagenet/ILSVRC/Data/CLS-LOC/

Negative Scouter for Imagenet dataset

python train.py --dataset ImageNet --model resnest26d --batch_size 70 --epochs 20 \
--num_classes 10 --use_slot true --use_pre false --loss_status -1 --slots_per_class 1 \
--power 2 --to_k_layer 3 --lambda_value 1 --vis false --channel 2048 --freeze_layers 0 \
--dataset_dir ../data/imagenet/ILSVRC/Data/CLS-LOC/

Visualization of Positive Scouter for Imagenet dataset

python test.py --dataset ImageNet --model resnest26d --batch_size 70 --epochs 20 \
--num_classes 10 --use_slot true --use_pre false --loss_status 1 --slots_per_class 1 \
--power 2 --to_k_layer 3 --lambda_value 1 --vis true --channel 2048 --freeze_layers 0 \
--dataset_dir ../data/imagenet/ILSVRC/Data/CLS-LOC/

Visualization of Negative Scouter for Imagenet dataset

python test.py --dataset ImageNet --model resnest26d --batch_size 70 --epochs 20 \
--num_classes 10 --use_slot true --use_pre false --loss_status -1 --slots_per_class 1 \
--power 2 --to_k_layer 3 --lambda_value 1 --vis true --channel 2048 --freeze_layers 0 \
--dataset_dir ../data/imagenet/ILSVRC/Data/CLS-LOC/

Visualization using torchcam for Imagenet dataset

python torchcam_vis.py --dataset ImageNet --model resnest26d --batch_size 70 \
--num_classes 10 --grad true --use_pre true \
--dataset_dir ../data/imagenet/ILSVRC/Data/CLS-LOC/ \
--grad_min_level 0

MNIST Dataset

Pre-training for MNIST dataset

python train.py --dataset MNIST --model resnet18 --batch_size 64 --epochs 10 \
--num_classes 10 --use_slot false --vis false --aug false

Positive Scouter for MNIST dataset

python train.py --dataset MNIST --model resnet18 --batch_size 64 --epochs 10 \
--num_classes 10 --use_slot true --use_pre true --loss_status 1 --slots_per_class 1 \
--power 1 --to_k_layer 1 --lambda_value 1. --vis false --channel 512 --aug false

Negative Scouter for MNIST dataset

python train.py --dataset MNIST --model resnet18 --batch_size 64 --epochs 10 \
--num_classes 10 --use_slot true --use_pre false --loss_status -1 --slots_per_class 2 \
--power 2 --to_k_layer 1 --lambda_value 1.5 --vis false --channel 512 --aug false --freeze_layers 3

Visualization of Positive Scouter for MNIST dataset

python test.py --dataset MNIST --model resnet18 --batch_size 64 --epochs 10 \
--num_classes 10 --use_slot true --use_pre true --loss_status 1 --slots_per_class 1 \
--power 1 --to_k_layer 1 --lambda_value 1. --vis true --channel 512 --aug false

Visualization of Negative Scouter for MNIST dataset

python test.py --dataset MNIST --model resnet18 --batch_size 64 --epochs 10 \
--num_classes 10 --use_slot true --use_pre false --loss_status -1 --slots_per_class 2 \
--power 2 --to_k_layer 1 --lambda_value 1.5 --vis true --channel 512 --aug false --freeze_layers 3

Visualization using torchcam for MNIST dataset

python torchcam_vis.py --dataset MNIST --model resnet18 --batch_size 64 \
--num_classes 10 --grad true --use_pre true

Con-Text Dataset

Pre-training for ConText dataset

python train.py --dataset ConText --model resnest26d --batch_size 200 --epochs 100 \
--num_classes 30 --use_slot false --vis false \
--dataset_dir ../data/con-text/JPEGImages/

Positive Scouter for ConText dataset

python train.py --dataset ConText --model resnest26d --batch_size 200 --epochs 100 \
--num_classes 30 --use_slot true --use_pre true --loss_status 1 --slots_per_class 3 \
--power 2 --to_k_layer 3 --lambda_value .2 --vis false --channel 2048 \
--dataset_dir ../data/con-text/JPEGImages/

Negative Scouter for ConText dataset

python train.py --dataset ConText --model resnest26d --batch_size 200 --epochs 100 \
--num_classes 30 --use_slot true --use_pre true --loss_status -1 --slots_per_class 3 \
--power 2 --to_k_layer 3 --lambda_value 1. --vis false --channel 2048 \
--dataset_dir ../data/con-text/JPEGImages/

Visualization of Positive Scouter for ConText dataset

python test.py --dataset ConText --model resnest26d --batch_size 200 --epochs 100 \
--num_classes 30 --use_slot true --use_pre true --loss_status 1 --slots_per_class 3 \
--power 2 --to_k_layer 3 --lambda_value 1. --vis true --channel 2048 \
--dataset_dir ../data/con-text/JPEGImages/

Visualization of Negative Scouter for ConText dataset

python test.py --dataset ConText --model resnest26d --batch_size 200 --epochs 100 \
--num_classes 30 --use_slot true --use_pre true --loss_status -1 --slots_per_class 3 \
--power 2 --to_k_layer 3 --lambda_value 1. --vis true --channel 2048 \
--dataset_dir ../data/con-text/JPEGImages/

Visualization using torchcam for ConText dataset

python torchcam_vis.py --dataset ConText --model resnest26d --batch_size 200 \
--num_classes 30 --grad true --use_pre true \
--dataset_dir ../data/con-text/JPEGImages/

CUB-200 Dataset

Pre-training for CUB-200 dataset

python train.py --dataset CUB200 --model resnest50d --batch_size 64 --epochs 150 \
--num_classes 25 --use_slot false --vis false --channel 2048 \
--dataset_dir ../data/bird_200/CUB_200_2011/CUB_200_2011/

Positive Scouter for CUB-200 dataset

python train.py --dataset CUB200 --model resnest50d --batch_size 64 --epochs 150 \
--num_classes 25 --use_slot true --use_pre true --loss_status 1 --slots_per_class 5 \
--power 2 --to_k_layer 3 --lambda_value 10 --vis false --channel 2048 --freeze_layers 2 \
--dataset_dir ../data/bird_200/CUB_200_2011/CUB_200_2011/

Negative Scouter for CUB-200 dataset

python train.py --dataset CUB200 --model resnest50d --batch_size 64 --epochs 150 \
--num_classes 25 --use_slot true --use_pre true --loss_status -1 --slots_per_class 3 \
--power 2 --to_k_layer 3 --lambda_value 1. --vis false --channel 2048 --freeze_layers 2 \
--dataset_dir ../data/bird_200/CUB_200_2011/CUB_200_2011/

Visualization of Positive Scouter for CUB-200 dataset

python test.py --dataset CUB200 --model resnest50d --batch_size 64 --epochs 150 \
--num_classes 25 --use_slot true --use_pre true --loss_status 1 --slots_per_class 5 \
--power 2 --to_k_layer 3 --lambda_value 10 --vis true --channel 2048 --freeze_layers 2 \
--dataset_dir ../data/bird_200/CUB_200_2011/CUB_200_2011/

Visualization of Negative Scouter for CUB-200 dataset

python test.py --dataset CUB200 --model resnest50d --batch_size 64 --epochs 150 \
--num_classes 25 --use_slot true --use_pre true --loss_status -1 --slots_per_class 3 \
--power 2 --to_k_layer 3 --lambda_value 1. --vis true --channel 2048 --freeze_layers 2 \
--dataset_dir ../data/bird_200/CUB_200_2011/CUB_200_2011/

Visualization using torchcam for CUB-200 dataset

python torchcam_vis.py --dataset CUB200 --model resnest50d --batch_size 150 \
--num_classes 25 --grad true --use_pre true \
--dataset_dir ../data/bird_200/CUB_200_2011/CUB_200_2011/

Acknowledgements

This work was supported by Council for Science, Technology and Innovation (CSTI), cross-ministerial Strategic Innovation Promotion Program (SIP), "Innovative AI Hospital System" (Funding Agency: National Institute of Biomedical Innovation, Health and Nutrition (NIBIOHN)).

Publication

If you want to use this work, please consider citing the following paper.

@inproceedings{li2021scouter,
 author = {Liangzhi Li and Bowen Wang and Manisha Verma and Yuta Nakashima and Ryo Kawasaki and Hajime Nagahara},
 booktitle = {IEEE International Conference on Computer Vision (ICCV)},
 pages = {},
 title = {SCOUTER: Slot Attention-based Classifier for Explainable Image Recognition},
 year = {2021}
}

SCOUTER: Slot Attention-based Classifier for Explainable Image Recognition

Related tags

Overview

SCOUTER: Slot Attention-based Classifier for Explainable Image Recognition

Abstract

Model Structure

Usage

Enable distributed training (if desired)

Imagenet

Training for Imagenet dataset (Base Model)

Positive Scouter for Imagenet dataset

Negative Scouter for Imagenet dataset

Visualization of Positive Scouter for Imagenet dataset

Visualization of Negative Scouter for Imagenet dataset

Visualization using torchcam for Imagenet dataset

MNIST Dataset

Pre-training for MNIST dataset

Positive Scouter for MNIST dataset

Negative Scouter for MNIST dataset

Visualization of Positive Scouter for MNIST dataset

Visualization of Negative Scouter for MNIST dataset

Visualization using torchcam for MNIST dataset

Con-Text Dataset

Pre-training for ConText dataset

Positive Scouter for ConText dataset

Negative Scouter for ConText dataset

Visualization of Positive Scouter for ConText dataset

Visualization of Negative Scouter for ConText dataset

Visualization using torchcam for ConText dataset

CUB-200 Dataset

Pre-training for CUB-200 dataset

Positive Scouter for CUB-200 dataset

Negative Scouter for CUB-200 dataset

Visualization of Positive Scouter for CUB-200 dataset

Visualization of Negative Scouter for CUB-200 dataset

Visualization using torchcam for CUB-200 dataset

Acknowledgements

Publication

Owner

scene-linear test images

Table Extraction Tool

Open Source Computer Vision Library

Hiiii this is the Spanish for Linux and win 10 and in the near future the english version of PortScan my new tool on which you can see what ports are Open only with the IP adress.

code for our ICCV 2021 paper "DeepCAD: A Deep Generative Network for Computer-Aided Design Models"

Convert scans of handwritten notes to beautiful, compact PDFs

Code release for Hu et al., Learning to Segment Every Thing. in CVPR, 2018.

computer vision, image processing and machine learning on the web browser or node.

A curated list of awesome synthetic data for text location and recognition

SceneCollisionNet This repo contains the code for "Object Rearrangement Using Learned Implicit Collision Functions", an ICRA 2021 paper. For more info

Corner-based Region Proposal Network

ARU-Net - Deep Learning Chinese Word Segment

Tools for manipulating and evaluating the hOCR format for representing multi-lingual OCR results by embedding them into HTML.

A bot that plays TFT using OCR. Keeps track of bench, board, items, and plays the user defined team comp.

Pixie - A full-featured 2D graphics library for Python

Um simples projeto para fazer o reconhecimento do captcha usado pelo jogo bombcrypto

A toolbox of scene text detection and recognition

Handwritten Text Recognition (HTR) system implemented with TensorFlow (TF) and trained on the IAM off-line HTR dataset. This Neural Network (NN) model recognizes the text contained in the images of segmented words.

MONAI Label is a server-client system that facilitates interactive medical image annotation by using AI.

Markup for note taking