Skip to content

aisu-programming/3D-aligned-2D-mouse-pose-estimation

Repository files navigation

Aligning 3D Latent Space for better 2D Mouse Pose Estimation

(Updated on 2024/12/06)

Utilized InfoNCE loss to align latent spaces from different-view 2D image inputs, allowing the model to learn a sense of 3D from 2D images.

Details

Datasets are stored in datasets.

Class objects for loading/augmenting images in datasets are stored in data.

Models such as Scalable UNet are stored in models.

Use download_datasets.sh to download images of MARS dataset.

Use build_resized_PE_dataset_from_local_YOLO.py to generate the cropped dataset for Top-down methods.

Use the scripts in scripts or train to train the pipeline (with hyperparameter search using wandb).