Semi-supervised Learning Model Weights for Detecting Floating Plastic Litter

Model weights The semi-supervised learning (SSL) models for plastic litter detection, are produced in the Experiment 1 of the following study: Jia, T., Taormina, R., de Vries, R., Kapelan, Z., van Emmerik, T. H., Vriend, P., & Okkerman, I. (2025). A semi-supervised learning-based framework for quantifying litter fluxes in river systems. Water Research, 124833. https://doi.org/10.1016/j.watres.2025.124833 The code for this publication is here: https://github.com/TianlongJia/deep_plastic_Flux_SSL The SSL models are developed in two stages: (1) pre-training and (2) fine-tuning. Details of the methodology are provided in the associated publication. The folder naming convention indicates: (1) the backbone architecture used during pre-training, (2) the number of unlabeled images used for pre-training, and (3) the number of pre-training epochs. For example, a ResNet50 backbone is pre-trained on 25k unlabeled images for 100 epochs. A Faster R-CNN model built on this pre-trained ResNet50 is then fine-tuned, validated, and tested using labeled images. The pre-trained model checkpoints and other output files are stored in the directory: SSL_RN50_25K_100e/pre_train/. The fine-tuned model checkpoints and other output files are stored in the directory: SSL_RN50_25K_100e/fine_tune/. During pre-training process, the model is first pre-trained using SwAV with the VISSL library. The resulting checkpoint is model_final_checkpoint_phase99.torch. This checkpoint is then converted into a format compatible with Detectron2. The converted weights are saved in the detectron2/ folder, and are then used for the fine-tuning process. During fine-tuning process, the model is fine-tuned on multiple datasets with different number of labeled images, such as SSL_10per/ folder contains fine-tuning scripts and weights of models, that is fine-tuned on 10% of labeled images. Additionally, some models are fine-tuned multiple times with different random seed. The results of these runs are saved in different folders if available, e.g., from SSL_100per_1/ to SSL_100per_10/ folder. Cite this paper If you find the code or publication is useful in your research or wish to refer to the paper, please use the following BibTeX entry. @article{jia2025semi, title={A semi-supervised learning-based framework for quantifying litter fluxes in river systems}, author={Jia, Tianlong and Taormina, Riccardo and de Vries, Rinze and Kapelan, Zoran and van Emmerik, Tim HM and Vriend, Paul and Okkerman, Imke}, journal={Water Research}, pages={124833}, year={2025}, publisher={Elsevier}}