Final Report Summary - VISCUL (Visual Culture for Image Understanding)
Traditionally, Computer Vision methods learn each visual concept separately, starting from scratch every time. The objective of this project was to break from this tradition by developing methods that allow computers to continuously learn visual concepts on top of what they already know, getting closer to how humans learn. A core issue is learning under weak supervision, in particular learning models that can localize objects in images, but without the need for expensive manual annotations indicating the location of the objects in the training images.
Some of the central outcomes of this project are:
- techniques for transferring visual knowledge from visual concepts which have location annotation for training, to classes that do not have them.
- techniques for training object class detectors with little human intervention. These schemes deliver high quality detectors and box annotations with substantially lower human annotation time than traditional drawing.
- several datasets with high quality annotations intended to push the community towards exploring new areas, namely: eye-tracking for learning computer vision models (POET dataset); contextual relations between object and background regions at a large-scale (COCO-Stuff dataset); modeling and learning articulated structures (TigDog dataset).
- pioneered the new area of learning object class models from video. The project explored exciting new avenues, including the highly challenging scenario of learning articulated object classes from weakly supervised video.
- several technical advances in semantic segmentation, object class detection, boundary detection, and semantic part detection.
Some of the central outcomes of this project are:
- techniques for transferring visual knowledge from visual concepts which have location annotation for training, to classes that do not have them.
- techniques for training object class detectors with little human intervention. These schemes deliver high quality detectors and box annotations with substantially lower human annotation time than traditional drawing.
- several datasets with high quality annotations intended to push the community towards exploring new areas, namely: eye-tracking for learning computer vision models (POET dataset); contextual relations between object and background regions at a large-scale (COCO-Stuff dataset); modeling and learning articulated structures (TigDog dataset).
- pioneered the new area of learning object class models from video. The project explored exciting new avenues, including the highly challenging scenario of learning articulated object classes from weakly supervised video.
- several technical advances in semantic segmentation, object class detection, boundary detection, and semantic part detection.