Evaluation Parameters

The following parameters are used during evaluation and inference of a trained LENS model. These options configure the dataset, reference/query pairings, and optional visual and metric-based evaluation tools.

LENS Evaluation Parameters

Argument	Type	Description
--dataset	str	Top-level folder name under lens/dataset/ that contains the dataset being used for reference and/or query.
--camera	str	Subfolder indicating the DVS camera used (e.g., davis128, speck).
--reference	str	Name of the reference experiment directory inside the dataset (used to load the trained model and evaluate against).
--reference_places	int	Number of reference images to include during evaluation.
--query	str	Name of the query experiment directory to be evaluated against the reference.
--query_places	int	Number of query images to use for evaluation.
--dims	list[int, int]	Size to resize input images to before feeding into the network.
--roi_dim	list[int, int]	Size of the input images for the network.
--sim_mat	flag	Generates a similarity matrix between reference and query images, useful for visual inspection of performance.
--matching	flag	Enables evaluation against a ground truth .npy file with binary match labels.
--GT_tolerance	int	Match is considered correct if it falls within this many indices of the true match (i.e., recall@tolerance).
--PR_curve	flag	Outputs a .json file containing precision-recall curve data and generates a corresponding plot.
--sad	flag	Runs SAD (Sum of Absolute Differences) matching instead of neural similarity, useful for baseline comparisons.
--nocuda	flag	Disables GPU usage for inference, forcing all operations to run on CPU.