研究目的
To propose a deep convolutional network based supervised coarse-to-fine algorithm for optical flow measurement that achieves comparable performance to previous methods with a smaller model size.
研究成果
The proposed compact coarse-to-fine optical flow prediction framework based on CNNs achieves comparable performance to state-of-the-art methods with a significantly smaller model size (98% smaller than FlowNet2.0), making it more practical for mobile applications. However, it has limitations in handling fast movements and occlusions, suggesting future work on stacking networks or ensemble models.
研究不足
The proposed framework struggles with fast movement of small objects and performs poorly on datasets with large motions and occlusions, such as KITTI. Data augmentation mechanisms were not used, which could be included in future work to improve robustness.
1：Experimental Design and Method Selection:
The study uses a coarse-to-fine strategy with a spatial pyramid network. It involves a dilated correlation layer for measuring correspondences between image pairs and flow decoders for optical flow estimation. The network is trained end-to-end using a multi-scale training loss.
2：Sample Selection and Data Sources:
Standard optical flow benchmark datasets are used: Flying Chairs (22,872 image pairs, split into training and test sets), MPI Sintel (1,041 image pairs, with clean and final versions, split as recommended), and KITTI 2012 (real road images, used only for testing).
3：List of Experimental Equipment and Materials:
Two NVIDIA Titan X GPUs for training and testing in data parallel mode. Software: Implemented in PyTorch deep learning platform.
4：Experimental Procedures and Operational Workflow:
The network architecture includes a spatial pyramid with 4 layers. At each layer, input images are downsampled, dilated correlation is computed, and optical flow is estimated using decoders. Training uses Adam optimizer with specific parameters, batch sizes, and learning rate schedules. Fine-tuning is performed on MPI Sintel datasets.
5：Data Analysis Methods:
Performance is evaluated using average endpoint error (EPE) as the metric. Results are compared with state-of-the-art methods on benchmark datasets.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
