We use the LaMem dataset for training. Please download LaMem.
We use pre-trained MoCo v2 weights as our initialization, please download it here.
Labels of both machine memorability scores and human memorability scores are provided in this repo.
Train a MachineMem predictor, run:
python main_predictor.py --inter_aug [path to lamem]
For HumanMem predictor, run:
python main_predictor.py --inter_aug --label_filename ./humanmem_scores.txt [path to lamem]
Please note that the folder with images should be a subfolder of your path folder.
python main_predictor.py -e --resume [path to predictor] [path to your dataset]
Results will be saved to ./test_result.txt.
If you only have few images to evaluate, you might try our demo at project page or hugging face.
We provide our pre-trained MachineMem/HumanMem predictors at this link.