Report, Paper (WeaSUL 2021 ICLR - Rejected)
Weakly Supervised Object Localization (WSOL) uses classification models trained only on image-level labels to localize objects. Although different WSOL techniques have shown results in improving object localization performance under weak supervision, limited studies have analyzed the role of dataset distribution on the performance of WSOL techniques. In this paper, we study the effects of intra-class variance in the dataset using selected WSOL techniques. We argue that intra-class variance is an important factor when applying WSOL techniques. We observe from our experiments that 2 baseline methods have a drop in performance when trained on datasets with limited intra-class variance. Furthermore, we show that artificially augmenting the intra-class variance improves the localization performance.
- Training a deep neural network as a classification model
- Only class level labels are available
- Trained classification model is used to localize objects
- Generate object level and pixel level labels during testing
- Class Activation Mapping: the confidence score of predicted class is projected back to the preceding
convolution layer to generate the class activation maps (CAMs)
- Gradient Class Activation Mapping architecture. Grad-CAM uses gradient flow passing through last
convolutional layer to generate weights
- Full report on research and results for various experiments are documented in MuthireddyVSSR-RnDReport
- Plausible publication and unique selling point is in paper
pip install -r requirements.txt
- Given in requirements file
- PASCAL VOC dataset
--root
---VOC2012
----Annotations
----ImageSets
----JPEGImages
----SegmentationClass
----SegmentationObject
- YCB dataset
--root
---train
----001_chips_can
-----masks
---- ....
----077_rubiks_cube
-----masks
---val
----001_chips_can
-----masks
---- ....
----077_rubiks_cube
-----masks
- RoboCup@Work dataset
--root
---images
----train
-----axis
----- ....
----- motor
----val
-----axis
----- ....
----- motor
python3 train.py --batch_size: BATCH SIZE (int),
--save_after: Save weights after every n epochs,
--num_epochs: number of epochs,
--dataset: Have option of three datsets VOC, YCB, at_work,
--dataset_path: Give the path to dataset,
--backbone: Have option of three backbones vgg16, resnet18, squeezenet1_1,
--experiment_number: Give unique identity to experiment
python3 classification_eval.py --batch_size: BATCH SIZE (int),
--dataset: Have option of three datsets VOC, YCB, at_work ,
--dataset_path: Give the path to dataset,
--backbone: Have option of three backbones vgg16, resnet18, squeezenet1_1,
--checkpoint_path: Path for trained checkpoint
Evaluation on localization metrics for different datasets is carried out differently due to very different structure of datasets.
- PASCAL VOC dataset
python3 localization_eval_VOC.py --wsol_method: select cam or gradCAM,
--dataset_path: Give the path to dataset,
--backbone: Have option of three backbones vgg16, resnet18, squeezenet1_1,
--checkpoint_path: Path for trained checkpoint
- YCB dataset
python3 localization_eval_YCB.py --wsol_method: select cam or gradCAM,
--dataset_path: Give the path to dataset,
--backbone: Have option of three backbones vgg16, resnet18, squeezenet1_1,
--checkpoint_path: Path for trained checkpoint,
--masks_path: Give the path to masks dataset
- RoboCup@Work dataset
python3 localization_eval_atwork.py --wsol_method: select cam or gradCAM,
--dataset_path: Give the path to dataset,
--backbone: Have option of three backbones vgg16, resnet18, squeezenet1_1,
--checkpoint_path: Path for trained checkpoint,
--masks_path: Give the path to masks dataset