Solving the Visual Question Answering (VQA) task is a step towards achieving human-like reasoning capability of the machines. This paper proposes an approach to learn multimodal feature representation with adversarial training. The purpose of the adversarial training allows the model to learn from standard fusion methods in an unsupervised manner. The discriminator model is equipped with a siamese combinatin of two standard fusion method namely multimodal compact bilinear pooling and multimodal tucker fusion. Output multimodal feature representation from generator is a resultant of graph convolutional operation. The resultant multimodal representation of the adversarial training allows the proposed model to infer the correct answers from open-ended natural language questions from the VQA 2.0 dataset. An overall accuracy of 69.86\% demonstrates the accuracy of the proposed model.

There are many task in surveillance monitoring such as object detection, person identification, activity and action recognition etc. Integrating variety of surveillance task through a multimodal interactive system will benefit real-life deployment, and will also support human operators. We first introduce a dataset which is first of its kind and named as Surveillance Video Question Answering (SVideoQA) dataset. The multi-camera surveillance monitoring aspect is considered through the multimodal context of Video Question Answering (VideoQA) in the SVideoQA dataset. This paper proposes a deep learning model where VideoQA task on the SVideoQA dataset is attempted to solved in a manner where memory-driven relationship among appearance and motion aspect of the video features are captured. At each level of the relational reasoning respective attentive parts of the context of the motion and appearance features are identified forwarded through frame level and clip level relational reasoning module. Also, respective memories are updated which are again forwarded to the memory-relation module to finally predict the answer word. The proposed memory-driven multilevel relational reasoning is made compatible with the surveillance monitoring task through the incorporation of multi-camera relation module, which is able to capture and reason over the relationships among the video feeds across multiple cameras. Experimental outcome exhibits that the proposed memory-driven multilevel relational reasoning perform significantly better on the open-ended VideoQA task compared to other state-of-the art systems. The proposed method achieves an accuracy of 57\% and 57.6\% respectively for the single-camera and multi-camera task of the SVideoQA dataset.