Using Weka: This week, I learned to use Weka software. This is a very easy tool to use. It implements several Machine Learning algorithms and one can simply upload training and test data and run any of the built-in Machine Learning models. The one I tried this week was Naive Bayes algorithm.
Text Detection and Recognition:
Rosetta: Facebook’s method explained in [4]. It is based on Faster R-CNN and CNN. Not open source.
Recommender System (Netflix Challenge): Most of this week was spent in solving the Netflix Challenge.
Paper Read: Door and Cabinet Recognition Using Convolutional Neural Nets and Real-Time Method for Handle Detection and Grasping
What is the difference between Bernoulli distribution and Binomial distribution? A Bernoulli random variable has two possible outcomes : 0 or 1. A binomial distribution is the sum of independent and identically distributed Bernoulli random variables.
How to train Detectron2 with custom dataset? Most of the week was spent training Detectron2 with the custom dataset. Need to create a separate tutorial for that
Forward-Backward Algorithm -used for predicting the next hidden state given the sequence of observations
Viterbi Algorithm – used for predicting the predicting the sequence of observations given the state
Baum Welch Algorithm – used for parameter estimation
Using Detectron2 to train on a custom dataset:
Prepare the dataset: If the dataset is already in the COCO format, then life becomes a little easier. My dataset is already in the COCO format.
Register the dataset: To register the dataset that is already in the COCO format, use the following function.
Paper Reading:
Learning to Generalize Kinematic Models to Novel Objects: This paper presents a new method for predicting the kinematics and current configuration of various articulated objects such as refrigerator, cabinet, microwave, drawer, toaster, and oven. This enables the robot to be able to interact with the object without having to learn the interaction through demonstration or other time-consuming techniques. According to the framework discussed, the kinematic models of six object classes mentioned above is learned separately during training. The paper uses three neural networks to learn mixture density network for each class. The neural networks learn mixing parameters, means and diagonal covariances for mixture of gaussians model for each class. Then, during inference an RGBD sensor data is given to the model as an input along with its class (which is recognized using Mask R-CNN). The model then estimates its geometry, kinematic model parameters and current configuration. This method was demonstrated with a real MOVO robot to open a microwave door. This paper releases the dataset that was used in their experiments. It is a synthetic dataset but very valuable for future research in this area. The paper does not share any code. It would be helpful to know what deep learning techniques they used for training to estimate the kinematics of various articulated objects.