Junsuk Choe1,3*, Seong Joon Oh2*, Seungho Lee1, Sanghyuk Chun3, Zeynep Akata4, Hyunjung Shim1

  • Equal contribution

1 School of Integrated Technology, Yonsei University
2 Clova AI Research, LINE Plus Corp. 3 Clova AI Research, NAVER Corp. 4 University of Tübingen

Weakly-supervised object localization (WSOL) has gained popularity over the last years for its promise to train localization models with only image-level labels. Since the seminal WSOL work of class activation mapping (CAM), the field has focused on how to expand the attention regions to cover objects more broadly and localize them better. However, these strategies rely on full localization supervision to validate hyperparameters and for model selection, which is in principle prohibited under the WSOL setup. In this paper, we argue that WSOL task is ill-posed with only image-level labels, and propose a new evaluation protocol where full supervision is limited to only a small held-out set not overlapping with the test set. We observe that, under our protocol, the five most recent WSOL methods have not made a major improvement over the CAM baseline. Moreover, we report that existing WSOL methods have not reached the few-shot learning baseline, where the full-supervision at validation time is used for model training instead. Based on our findings, we discuss some future directions for WSOL.

Overview of WSOL performances 2016-2019. Above image shows that recent improvements in WSOL are illusory due to (1) different amount of implicit full supervision through validation and (2) a fixed score-map threshold to generate object boxes. Under our evaluation protocol with the same validation set sizes and oracle threshold for each method, CAM is still the best. In fact, our few-shot learning baseline, i.e., using the validation supervision (10 samples/class) at training time, outperforms existing WSOL methods.

1. Our dataset contribution

WSOL is an ill-posed problem when only image-level labels are available (see paper for an argument).
To be able to solve the WSOL task, certain amount of full supervision is inevitable, and prior WSOL approaches have utilized
different amount of implicit and explicit full supervision (usually through validation).
We propose to use a fixed amount of full supervision per method by carefully designing validation splits (called train-fullsup in the paper), such that different methods use the same amount of localization-labelled validation split.

In this section, we explain how each dataset is split, and introduce our data contributions (image collections and new annotations) on the way.

The dataset splits

split ImageNet CUB OpenImages
train-weaksup ImageNet "train" CUB-200-2011 "train" OpenImages30k "train" we_curated data
train-fullsup ImageNetV2 we_collected annotations CUBV2 we_collected images_and_annotations OpenImages30k "val" we_curated data
test ImageNet "val" CUB-200-2011 "test" OpenImages30k "test" we_curated data

We propose three disjoint splits for every dataset: train-weaksup,
train-fullsup, and test. The train-weaksup contains images with weak
supervision (the image-level labels). The train-fullsup contains images with
full supervision (either bounding box or binary mask). It is left as freedom for
the user to utilize it for hyperparameter search, model selection, ablative
studies, or even model fitting. The test split contains images with full
supervision; it must be used only for the final performance report. For
example, checking the test results multiple times with different model
configurations violates the protocol as the learner implicitly uses more full
supervision than allowed. The splits and their roles are more extensively explained
in the paper.

  • ImageNet
    • "train" and "val" splits of original ImageNet
      are treated as our train-weaksup and test.
    • ImageNetV2 is treated as our
      train-fullsup. Note that we have annotated bounding boxes on ImageNetV2.
  • CUB
    • "train" and "test" splits of original
      CUB-200-2011 are
      treated as our train-weaksup and test.
    • We contribute images and annotations that are similar as the original CUB,
      namely CUBV2.
  • OpenImages
    • We curate the existing OpenImagesV5
      for the task of WSOL.
    • We have randomly selected images from the original "train", "val", and
      "test" splits of the instance segmentation subset.

2. Dataset downloading and license

For original ImageNet and CUB datasets, please follow the common procedure to
download the datasets. In this section, we only explain how to obtain the less used
(or never used before) datasets. We also provide the license status for each dataset.
This section is for those who are interested in the full data for each dataset.
If the aim is to utilize the data for WSOL evaluation and/or training, please follow the links below:


Download images

We utilize 10,000 images in the Threshold0.7 split of
ImageNetV2 for our train-fullsup
split. We have annotated bounding boxes on those images.
Box labels exist in here and are licensed by NAVERCorp. under
Attribution 2.0 Generic (CC-BY-2.0).


Download images

We have collected and annotated CUBV2 on our own as the train-fullsup split.
We have ensured that the data distribution
follows the original CUB dataset and there is no duplicate image.
We have collected 5 images per class
(1,000 images total) from Flickr.
Box labels and license files of all images exist in here.
Both class and box labels are licensed by NAVERCorp under
Attribution 2.0 Generic (CC-BY-2.0).


Download images
Download segmentation masks

The WSOL community has relied on ImageNet and CUB datasets at least for the last three years.
It is perhaps time for us to move on. We provide a WSOL benchmark based on the OpenImages30k
dataset to provide a new perspective on the generalizability of
WSOL methods in the past and future. To make it suitable for the WSOL task,
we use 100 classes to ensure the minimum number of single-class samples for
each class. We have randomly selected 29,819, 2,500, and 5,000 images from the
original "train", "val", and "test" splits of
Corresponding metadata can be found in here.
The annotations are licensed by Google LLC under
Attribution 4.0 International (CC-BY-4.0).
The images are listed as having a
Attribution 2.0 Generic (CC-BY-2.0).

Dataset statistics

Below tables summarizes dataset statistics of each split.

# images/classes ImageNet 1,000 classes CUB 200 classes OpenImages 100 classes
train-weaksup ~1,200 ~30 ~300
train-fullsup 10 ~5 25
test 10 ~29 50


The licenses corresponding to our dataset contribution are summarized as follows

Dataset Images Class Annotations Localization Annotations
ImageNetV2 See the original Github See the original Github CC-BY-2.0 NaverCorp.
CUBV2 Follows original image licenses. See here. CC-BY-2.0 NaverCorp. CC-BY-2.0 NaverCorp.
OpenImages CC-BY-2.0 (Follows original image licenses. See here) CC-BY-4.0 Google LLC CC-BY-4.0 Google LLC

Detailed license files are summarized in the release directory.

Note: At the time of collection, images were marked as being licensed under
the following licenses:

Attribution-NonCommercial License
Attribution License
Public Domain Dedication (CC0)
Public Domain Mark

However, we make no representations or warranties regarding the license status
of each image. You should verify the license for each image yourself.

3. Code dependencies

Both the evaluation-only and eval+train scripts require only the following libraries:

pip freeze returns the version information as below:


4. WSOL evaluation

We support evaluation of weakly-supervised object localization (WSOL) methods
on CUB, ImageNet, and OpenImages. The main script for evaluation is We will show how to download the train-fullsup
(validation) and test set images and localization annotations.
An example evaluation script will be provided.

Prepare evaluation data

WSOL evaluation data consist of images and corresponding localization ground
truths. On CUB and ImageNet, they are given as boxes, and on OpenImages, they
are given as binary masks.

To prepare evaluation data, first, download ImageNet "val" split from
here and put the downloaded file on

Then, run the following command


The script will download the train-fullsup (validation) and test images at
dataset. Metadata and box annotations already exist in this repository
under metadata. OpenImages mask annotations are also downloaded by
the above script, and will be saved under dataset with the images.

The structure of image files looks like

    └── val2
        └── 0
            ├── 0.jpeg
            ├── 1.jpeg
            └── ...
        └── 1
        └── ...
    └── val
        ├── ILSVRC2012_val_00000001.JPEG
        ├── ILSVRC2012_val_00000002.JPEG
        └── ...
└── CUB
    └── 001.Black_footed_Albatross
        ├── Black_Footed_Albatross_0046_18.jpg
        ├── Black_Footed_Albatross_0002_55.jpg
        └── ...
    └── 002.Laysan_Albatross
    └── ...
└── OpenImages
    └── val
        └── 0bt_c3
            ├── 1cd9ac0169ec7df0.jpg
            ├── 1cd9ac0169ec7df0_ignore.png
            ├── 1cd9ac0169ec7df0_m0bt_c3_6932e993.png
            └── ...
        └── 0bt9lr
        └── ...
    └── test   
        └── 0bt_c3
            ├── 0a51958fcd523ae4.jpg
            ├── 0a51958fcd523ae4_ignore.png
            ├── 0a51958fcd523ae4_m0bt_c3_41344f12.png
            ├── 0a51958fcd523ae4_m0bt_c3_48f37c0f.png
            └── ...
        └── 0bt9lr
        └── ...

Prepare heatmaps to evaluate

Our WSOL evaluation evaluates heatmaps of the same width and height as the input
images. The evaluation script requires the heatmaps to meet the following

  1. Heatmap file structure.
  • Heatmaps shall be located at the user-defined <heatmap_root>.
  • <heatmap_root> folder contains the heatmap files with the file names dictated by the metadata/<dataset>/<split>/image_ids.txt files.
  • If an image_id has slashes (/), e.g. val2/995/0.jpeg, then the corresponding heatmaps shall be located at the corresponding sub-directories, e.g. <heatmap_root>/val2/995/0.npy.
  1. Heatmap data type.
  • Each heatmap file should be a .npy file that can be loaded as a numpy array with numpy.load().
  • The array shall be two-dimensional array of shape (height, width), same as the input image sizes.
  • The array shall be of type np.float.
  • The array values must be between 0 and 1.

Evaluate your heatmaps

We support three datasets, CUB, ImageNet, and OpenImages.

On CUB and ImageNet, we evaluate the MaxBoxAcc, the maximal box accuracy at
the optimal heatmap threshold, where the box accuracy is measured by the ratio
of images where the box generated from the heatmap overlaps with the ground
truth box with IoU at least 0.5. Please see the code and paper for the full

On OpenImages, we evaluate the PxAP, pixel average precision. We generate
the pixel-wise precision-recall curve, and compute the area under the curve.
Please see the code and paper for the full details.

We present an example call to the evaluation API below:

python --scoremap_root=train_log/scoremaps/ \
                     --metadata_root=metadata/ \
                     --mask_root=dataset/ \
                     --dataset_name=CUB \
                     --split=val \

When CUB evaluation data are downloaded at dataset using our download script
above, and the corresponding heatmaps are saved under train_log/scoremaps/,
then the MaxBoxAcc will be evaluated as a result of this call.

Testing the evaluation code

The test code for the evaluation modules is given at The unit tests ensure the correctness
of the evaluation logic, and potentially prevents unnoticed changes in the
functionalities of underlying libraries (e.g. OpenCV, Numpy).
To run the unit test, run


pip3 install nose may be required to install nose.

5. Library of WSOL methods

We support the training and evaluation of the following weakly-supervised object
localization (WSOL) methods. Our implementation of the methods can be found in
the wsol folder. Please add your own WSOL method in the list by making
a pull request.

We provide the full training and evaluation scripts on the provided WSOL methods.
Details will be explained in the next section.

Method Paper Original code
Class-Activation Mapping (CAM) CVPR'16 Code
Hide-and-Seek (HaS) ICCV'17 Code
Adversarial Complementary Learning (ACoL) CVPR'18 Code
Self-Produced Guidance (SPG) ECCV'18 Code
Attention-based Dropout Layer (ADL) CVPR'19 Code
CutMix ICCV'19 Code

6. WSOL training and evaluation

We describe the data preparation and training scripts for the above six prior
WSOL methods.

Prepare train+eval datasets

Our repository enables evaluation and training of WSOL methods on two
commonly-used benchmarks, CUB and ImageNet, and our newly-introduced benchmark
OpenImages. We describe below how to prepare those datasets.


Both the original ImageNet and
ImageNetV2 are required for WSOL
training. Note that "val" split of the original ImageNet is considered as test
split, and ImageNetV2 is used for split (train-fullsup) in our framework.

To prepare ImageNet data, download ImageNet "train" and "val" splits from
here and put the downloaded file on
dataset/ILSVRC2012_img_train.tar and dataset/ILSVRC2012_img_val.tar.

Then, run the following command on root directory to extract the images.


apt-get install parallel may be required to install parallel.

The structure of image files looks like

    └── train
        └── n01440764
            ├── n01440764_10026.JPEG
            ├── n01440764_10027.JPEG
            └── ...
        └── n01443537
        └── ...
    └── val2
        └── 0
            ├── 0.jpeg
            ├── 1.jpeg
            └── ...
        └── 1
        └── ...
    └── val
        ├── ILSVRC2012_val_00000001.JPEG
        ├── ILSVRC2012_val_00000002.JPEG
        └── ...

Corresponding annotation files can be found in here.


Both the original
and our CUBV2 datasets are required for WSOL training. Note that CUBV2 is
considered as a validation split (train-fullsup). Then, run the following
command to download original CUB dataset and extract the image files on root


Note: you can also download the CUBV2 dataset from
here. Put
the downloaded file on dataset/CUBV2.tar directory and then run the above script.

The structure of image files looks like

└── CUB
    └── 001.Black_footed_Albatross
        ├── Black_Footed_Albatross_0001_796111.jpg
        ├── Black_Footed_Albatross_0002_55.jpg
        └── ...
    └── 002.Laysan_Albatross
    └── ...

Corresponding annotation files can be found in here.


We provide a new WSOL benchmark, OpenImages30k,
based on OpenImagesV5.

To download and extract files, run the following command on root directory


Note: you can also download the OpenImages30k dataset from here
, masks).
Put the downloaded and
files in dataset directory and run the above script.

The structure of image files looks like:

└── OpenImages
    └── train
        └── 0bt_c3
            ├── 0a9b7df4d832baf7.jpg
            ├── 0abee225b2418fe7.jpg
            └── ...
        └── 0bt9lr
        └── ...
    └── val
        └── 0bt_c3
            ├── 1cd9ac0169ec7df0.jpg
            ├── 1cd9ac0169ec7df0_ignore.png
            ├── 1cd9ac0169ec7df0_m0bt_c3_6932e993.png
            └── ...
        └── 0bt9lr
        └── ...
    └── test   
        └── 0bt_c3
            ├── 0a51958fcd523ae4.jpg
            ├── 0a51958fcd523ae4_ignore.png
            ├── 0a51958fcd523ae4_m0bt_c3_41344f12.png
            ├── 0a51958fcd523ae4_m0bt_c3_48f37c0f.png
            └── ...
        └── 0bt9lr
        └── ...

Corresponding annotation files can be found in here.

Run train+eval

We support the following architecture and method combinations:

  • Architectures.

    • vgg16
    • inception_v3
    • resnet50
  • Methods (see Library of WSOL methods and paper for descriptions).

    • cam
    • has
    • acol
    • spg
    • adl
    • cutmix

Below is an example command line for the train+eval script.

python --dataset_name OpenImages \
               --architecture vgg16 \
               --wsol_method cam \
               --experiment_name OpenImages_vgg16_CAM \
               --pretrained TRUE \
               --num_val_sample_per_class 5 \
               --large_feature_map FALSE \
               --batch_size 32 \
               --epochs 10 \
               --lr 0.00227913316 \
               --lr_decay_frequency 3 \
               --weight_decay 5.00E-04 \
               --override_cache FALSE \
               --workers 4

See for the full descriptions of the arguments, especially the method-specific hyperparameters.

7. Code license

This project is distributed under MIT license.

Copyright (c) 2020-present NAVER Corp.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

8. How to cite

  title={Evaluating Weakly Supervised Object Localization Methods Right},
  author={Choe, Junsuk and Oh, Seong Joon and Lee, Seungho and Chun, Sanghyuk and Akata, Zeynep and Shim, Hyunjung},
  journal={arXiv preprint arXiv:2001.07437},