MASON proposes a simple, yet very effective method to localize dominant foreground objects in an image, to pixel-level precision. The proposed method `MASON’ (Model-AgnoStic ObjectNess) uses a deep convolutional network to generate category-independent and model-agnostic heat maps for any image. The network is not explicitly trained for the task and hence, can be used off-the-shelf in tandem with any other network or task. We show that this framework scales to a wide variety of images, and illustrate the effectiveness of MASON in three varied application contexts.
Task 1: Foreground Segmentation
The effectiveness of MASON in fine-grained object localization is captured in the figure (Column 3). The proposed method (Column 3) is more effective than just using Grabcut (Column 4) in foreground segmentation.
Task 2: Extending object detectors to perform instance segmentations
Task 3: Improving the quality of detection datasets
The picture captures the effectiveness of \method to enhance bad annotations in a dataset. False positive annotations are removed and remaining annotations are made tighter by MASON. The image is the 4005th frame from Video 0 of `Bookstore’ scene in the Stanford Drone Dataset. Red color bounding boxes annotate Pedestrians while blue color annotate Biker.
Code for MASON: https://github.com/JosephKJ/MASON