Hyperparameter Optimization for predict_tile

[mtaniguchiking]

Hello! I'm evaluating DeepForest's predict_tile as a CV tree detection method for OFO and would appreciate insight/feedback on my current approach:

As a starting point for the evaluation, I'm optimizing predict_tile's hyperparameters on a forest orthomosaic to gauge the effectiveness of predict_tile and to get a sense for how the performance (i.e. f-score) of predict_tile changes depending on its hyperparameters. Before considering multiple parameters, I began by exploring the influence of individual parameters - as an example, here is a preliminary result from a range of patch_size values:

My current approach for hyperparameter fine-tuning involves running predict_tile on an orthomosaic with different hyperparameter combinations before comparing the DeepForest tree detections against a ground-measured reference tree map.

My plan involves performing the following searches for hyperparameter fine-tuning:
Random Search with NMS (use_soft_nms = False):

• Parameters to optimize: patch_size, patch_overlap, and iou_threshold, orthomosaic resolution (image rescale factor)

I’m curious about the role of thresh when use_soft_nms = False. Documentation indicates thresh only filters bboxes after soft NMS is performed (therefore, it is not relevant when use_soft_nms = False), but I would like to confirm this. I’m also more generally interested in other considerations (specific to DeepForest) I should take into account as I optimize these parameters.

[bw4sz]

Closing, since was responded by email. #471 opened issue.