Unsupervised Learning of Depth and Ego-Motion from Monocular Video Using 3D Geometric Constraints

Reza Mahjourian, Martin Wicke, Anelia Angelova
CVPR 2018. [PDF] [Website] [Google I/O ‘18]

Learning depth and ego-motion (camera motion) unsupervised from monocular (single-camera) video. The apparent motion of pixels between adjacent frames contains information on scene depth and camera motion. This method uses reprojections based on predicted depth and ego-motion as supervisory signals. Introduced a new TensorFlow op to align point clouds from adjacent frames, allowing us to directly learn the 3D scene geometry.

We evaluate the method by training models on a low-quality uncalibrated video dataset recorded on a phone camera while riding a bicycle. The trained model is used to predict depth on KITTI, ranking among top performing prior methods which are trained on KITTI itself.

Reza Mahjourian, Risto Miikkulainen
[PhD Proposal]

Discovering suitable cost functions allowing Guided Policy Search (GPS) to solve tasks that require planning for intermediate goals. As the animation above shows, direct optimization may lead to local optima.

Geometry-Based Next Frame Prediction from Monocular Video

Reza Mahjourian, Martin Wicke, Anelia Angelova
IEEE Intelligent Vehicles 2017. [PDF]

We use the scene geometry to make more accurate next-frame predictions. A recurrent neural network using convolutional LSTM cells is trained to predict depth from a sequence of monocular video frames. The depth prediction along with the current video image and the camera trajectory is then used to compute a prediction for the next frame.

Neuroevolutionary Planning for Robotic Control

Reza Mahjourian, Risto Miikkulainen
BEACON 2014 [PDF] [PhD Proposal]

Among our key findings is that moderate amounts of actuation noise improve the controller’s robustness against a joint malfunction. Interestingly, actuation noise improves the expected fitness and reduces variance even under normal conditions in absense of any malfunction.

With proper input and output setup, evolution is able to discover controllers with precise behavior. However, when extending the task to require strategy and planning, finding solutions becomes hard. This work proposes a new evolutionary method to discover and complete subtasks leading to completion of the original goal.

An Evolutionary Feature Discovery Method for Reinforcement Learning

Reza Mahjourian, Peter Stone
GECCO 2013 submission. [PDF] [Code]

We present an evolutionary algorithm for generating and evaluating candidate feature sets for learning a task using gradient descent Sarsa(λ) as a linear method.

Studying Impact of Domain Ergodicity and Stochasticity on Reinforcement Learning with Self-Play

Reza Mahjourian, Prateek Maheshwari, Risto Miikkulainen, 2011
[Code] [Proposal] [Report]

This work studies different hypotheses on why TD learning worked so well for backgammon. Does backgammon have particular attributes that make it suitable for reinforcement learning? Can they be exploited to design better learning algorithms for other domains? We ran experiments that showed stochasticity to be a major facilitator for reinforcement learning with self-play.

Active Learning for Robotic Tasks

Reza Mahjourian, Peter Stone, 2011
[Code] [Project Report]

Used an ensemble of neural networks and selected samples by prioritizing inputs where the networks in the ensemble disagreed the most (most variance in predictions).

An Architectural Style for Data-Driven Systems

Reza Mahjourian
International Conference on Software Reuse (ICSR), 2008
[Code] [PDF] [Extended]

We present XPage, an architectural style which is especially designed for building data-driven systems. This architectural style is the basis for an open-source light-weight web framework, which transparently translates high-level page specs into server-side scripts on request.

An Approximation Algorithm for Conflict-Aware Broadcast Scheduling in Wireless Ad Hoc Networks

Reza Mahjourian, Feng Chen, Ravi Itwari, My Thai, Hongqiang Zhai, Yuguang Fang
The ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc), 2008 [PDF]

We introduce and prove correctness of a constant approximation algorithm for minimum-latency conflict-aware broadcast scheduling in wireless networks. A constant approximation algorithm is a polynomial-time solution to an NP-hard problem whose solution is within a constant multiple of the optimal solution.

Software Connector Classification and Selection for Data-Intensive Systems

Chris A. Mattmann, David Woollard, Nenad Medvidovic, and Reza Mahjourian International Workshop on Incorporating COTS Software into Software Systems, 2007 [PDF]

Data-intensive systems and applications transfer large volumes of data and metadata to highly distributed users separated by geographic distance and organizational boundaries. An influential element in these large volume data transfers is the selection of the appropriate software connector that satisfies user constraints on the required data distribution scenarios. In this paper we present a systematic approach for selecting software connectors based on eight key dimensions of data distribution that we use to represent the data distribution scenarios.

Reza Mahjourian

PhD candidate at UT Austin. Student researcher at Google Brain Robotics.

Reza Mahjourian