Most of the computer and robot-vision algorithms are designed mainly for opaque objects and non-opaque objects have received less attention, in spite of them being omnipresent in man-made environments. With an increasing usage of such objects, especially those made of glass, plastic etc., it becomes necessarily important to detect this class of objects while building a robot navigation system. Obstacle avoidance forms a primary yet challenging task in mobile robot navigation. The main objective of this paper is to present an algorithm to detect and avoid obstacles that are made of semi-transparent materials, such as plastic or glass. The algorithm makes use of a technique called the collective-reward based approach to detect such objects from single images captured by an uncalibrated camera in a live video stream. Random selection techniques are incorporated in the method to make the algorithm run in real-time. A mobile robot then uses the information after detection to perform an obstacle avoidance maneuver. Experiments were conducted on a real robot to test the efficacy of the algorithm.