We have created a dataset of frames extracted from videos of Irish Sign Language (ISL) for sign language recognition. The dataset was collected by recording human subjects executing ISL hand-shapes and movements. Frames were extracted from the videos producing a total of 52,688 images for the 23 static common hand-shapes. Given that some of the frames were relativity similar we designed a new method for removing redundant frames based on labelling the hand images by using axis of least inertia - Hand Orientation Redundancy Filter (HORF) - and we compare the results with an iterative method - Iterative Redundancy Filter (IRF). This selection process method selects the most different images in order to keep the dataset diverse. The IRF dataset contains 50,000 images whereas the HORF consists of 27,683 images. Finally, we tested two classifiers over the HORF dataset and compared the results with the IRF dataset