DV-491 ⁃ Investigate/research classification architectures for IDS use-case

[sync-by-unito[bot]]

┆Issue is synchronized with this Jira Task by Unito

[sync-by-unito[bot]]

➤ Thabied Majal commented:

Decision: utilize OCR > Classification process. This process is the standard used in CV production applications and applies to the Panthera use-case. Although this process has a longer inference time, the growing variation in background vegetation and varying image-volume species distribution, it calls for localization within the image. The goal is also to use less images per specie to achieve a robust classifier

[sync-by-unito[bot]]

➤ Thabied Majal commented:

Currently the best results are coming from YOLO > ViT pipelines. YOLO only passes over images once and yields good results as well as fast inference times. The concern is that ViT models are complex and require large compute. They do run on CPU’s, which all Panthera users will be using when running the classifier, but more experimentation is required with multiple different models and dataset structures to investigate best use-case for IDS implementation