Winter 2021 Project CS230
ABSTRACT
Post-earthquake damage surveys currently require teams of domain experts to visually inspect buildings to determine their safety, which is slow and subjective. Efforts to automate the process using computer vision have been limited due to the time and resource cost of labeling earthquake survey images. In this project I use pseudo-labeling to take advantage of large numbers of unlabeled reconnaissance images available on the web in a semi-supervised learning approach. I investigate the effects of changing the start-epoch of pseudo-labeling during training. Results show that prediction accuracy improved by up to 5% on a test set from the unlabeled data, although improvement is sensitive to the accuracy of the base model.
PAPER
http://cs230.stanford.edu/projects_winter_2021/reports/70759638.pdf
DATASET
https://apps.peer.berkeley.edu/phi-net/
Gao Y., & Mosalam K.M. (2018). Deep Transfer Learning for Image-based Structural Damage Recognition, Computer-Aided Civil and Infrastructure Engineering, 33(9): 748-768.
Gao, Y., & Mosalam, K. M. (2019). PEER Hub ImageNet (Φ-Net): A Large-Scale Multi-Attribute Benchmark Dataset of Structural Images. PEER Report No.2019/07, Pacific Earthquake Engineering Research Center, University of California, Berkeley, CA.
Gao, Y., & Mosalam, K. M. (2020). PEER Hub ImageNet: A Large-Scale Multiattribute Benchmark Data Set of Structural Images. Journal of Structural Engineering, 146(10), 04020198.
(Usage requires Tasks 1 and 2)
COMMAND LINE ARGUMENTS
--task: Task 1: Scene Level (Pixel, Object, Structure), Task 2: Damage State (Damaged, Undamaged)
--semisupervised: True/False use unlabeled images for pseudo-label training
--path: Location of folders containing datasets for each task
--val_split: Proportion of (labeled) training set to use as validation
--batch_size: Number of examples per training batch
--epochs: Number of epochs to train from data
--lr: Adam optimizer learning rate
--alpha_range: List of length two defining the epoch to start including pseudo-labels in loss function, and the epoch to end the ramp-up of loss weighting
--crop_dataset: Decrease the size of the training dataset for debugging speed
USAGE EXAMPLES
Supervised: python main.py --task 1 --epochs 20
Semisupervised: python main.py --task 1 --semisupervised --alpha_range 3 5 --epochs 5