IPAB Workshop

Speaker: Holger Caesar

Talk title: Joint Calibration for Semantic Segmentation

Talk abstract: Semantic segmentation is the task of assigning a class-label to each pixel in an image. We propose a region-based semantic segmentation framework which handles both full and weak supervision, and addresses three common problems:

 (1) Objects occur at multiple scales and therefore we should use regions at multiple scales. However, these regions are overlapping which creates conflicting class predictions at the pixel level.

 (2) Class frequencies are highly imbalanced in realistic datasets.

 (3) Each pixel can only be assigned to a single class, which creates competition between classes. We address all three problems with a joint calibration method which optimizes a multi-class loss defined over the final pixel-level output labeling, as opposed to simply region classification.Results on the SIFT Flow dataset show that in the fully supervised setting our method yields state-of-the-art results, while it improves over the state-of-the-art by 9% in the weakly supervised setting.

 

Speaker: Dimitrios Papadopoulos

Talk Title: Deep Multiple Instance Learning with quality control for weakly supervised object localization

Talk Abstract: We address weakly supervised object localization (WSOL) in the Multiple Instance Learning (MIL) framework, which typically iteratively alternate two steps: selecting windows covering the object in positive training images, and updating a window classifier given the current selection. Recent, state-of-the-art WSOL methods build on Convolutional Neural Nets (CNN) pre-trained from a large external dataset. However, they use CNNs as a static feature extraction mechanism: during MIL only a simple SVM classifier is updated. We propose to perform full deep learning and update the complete CNN during MIL, including the lower feature representation layers. In order to reduce the impact of errors in the current selection of positive examples, we also propose quality control mechanisms to determine which examples are used to update the CNN. We demonstrate the benefits of both contributions through experiments on the popular PASCAL VOC 2007 dataset.

Speaker: Stavros Gerakaris

Talk title: Best Response Strategies Between Adaptive Agents with Censored Observations

Talk abstract: We are considering an emerging problem in the domain of ad display internet auctions (Ad Exchange), where a publisher is deciding his best response strategy against an advertiser’s unknown buying strategy, through continuous interactions with censored observations. We model it as a Bayesian Stochastic Game of limited information and we are making the critical assumptions that the marketplace consists of a single buyer and a single seller, the advertiser chooses from a finite and well defined set of strategies and we are within a finite time horizon repeated auction setting. In this talk I will present experiments and results on this problem, by adapting methods of agent coordination (HBA algorithm) and distribution approximation (KM estimator).