Abstract
Learning from sequential data as it becomes available opens up a diverse set of applications that would otherwise require large datasets to accurately represent rapidly changing environments, such as those found in construction environments. Due to labor shortages and hard manual labor, the construction industry is increasingly looking for automated solutions to increase efficiency. In this work, we present a model-agnostic online learning pipeline that utilizes the parameter estimation method RANSAC to generate ground truth labels which are used to iteratively update a deep learning model. The pipeline includes a dynamic dataset updated based on the amount of potential information to be learned. Specifically, each entry in the dataset is sampled with a probability inversely proportional to the model performance on those entries. Additionally, we utilize this same method to replace out-of-date entries as more data becomes available, ensuring a max size of the dataset with high learning potential. Each data entry in the dataset includes images with a temporal distance δ into the past. The images are stacked channel-wise to factor in the implicit temporal property of sequential video streams. Our experiments show that even with only retaining a modest dataset size this sampling strategy performs up to 15% better than a standard first-in-first-out dataset with a uniform sampling strategy.
Original language | English |
---|---|
Title of host publication | Proceedings of 2023 IEEE International Conference on Imaging Systems and Techniques |
Number of pages | 6 |
Publisher | IEEE |
Publication date | 19 Oct 2023 |
Article number | 10355699 |
ISBN (Print) | 979-8-3503-3084-7 |
DOIs | |
Publication status | Published - 19 Oct 2023 |
Event | 2023 IEEE International Conference on Imaging Systems and Techniques - Technical University of Denmark, Copenhagen, Denmark Duration: 17 Oct 2023 → 19 Oct 2023 |
Conference
Conference | 2023 IEEE International Conference on Imaging Systems and Techniques |
---|---|
Location | Technical University of Denmark |
Country/Territory | Denmark |
City | Copenhagen |
Period | 17/10/2023 → 19/10/2023 |
Keywords
- Training
- Robot vision systems
- Pipelines
- Video sequences
- Semantics
- Streaming media
- Cameras