Exploring Video Frame Redundancies for Efficient Data Sampling and Annotation in Instance Segmentation

In recent years, deep neural network architectures and learning algorithms have greatly improved the performance of computer vision tasks. However, acquiring and annotating large-scale datasets for training such models can be expensive. In this work, we explore the potential of reducing dataset sizes by leveraging redundancies in video frames, specifically for instance segmentation. To accomplish this, we investigate two sampling strategies for extracting keyframes, uniform frame sampling with adjusted stride (UFS) and adaptive frame sampling (AFS), which employs visual (Optical flow, SSIM) or semantic (feature representations) dissimilarities measured by learning free methods. In addition, we show that a simple copy-paste augmentation can bridge the big mAP gap caused by frame reduction. We train and evaluate Mask R-CNN with the BDD100K MOTS dataset and verify the potential of reducing training data by extracting keyframes in the video. With only 20% of the data, we achieve similar performance to the full dataset mAP; with only 33% of the data, we surpass it. Lastly, based on our findings, we offer practical solutions for developing effective sampling methods and data annotation strategies for instance segmentation models. Supplementary on https://github.com/jihun-yoon/EVFR.