Generating embeddings for new training samples

[KyriaAnnwyn]

How to generate embeddings for my training set?

[hukkelas]

Code for generating embeddings for new training datasets is currently not supported, but it should not be that much required.

The following code detects an image, crops the different detections and post processes the embedding/image to generate a single annotated image for each detection.

import torchvision
from fba.anonymizer.cse import CSEDetector
from fba.anonymizer.post_process import process_cse_detections

input_path = "coco_val2017_000000001000.jpg"
im = torchvision.io.read_image(input_path, mode=torchvision.io.ImageReadMode.RGB)
detector = CSEDetector()
detections = detector(im)
embed_map = detections["embed_map"]
if detections is None:
    exit()
post_process_cfg = dict(
    target_imsize=(288, 160),
    exp_bbox_cfg=dict(percentage_background=0.3, axis_minimum_expansion=.1),
    exp_bbox_filter=dict(minimum_area=32*32, min_bbox_ratio_inside=0, aspect_ratio_range=[0, 99999]),
)
detections = process_cse_detections(
    im, **detections, **post_process_cfg
)
for i, sample in enumerate(detections):
    print("Detection:", i)
    for key, value in sample.items():
        if key == "exp_bbox":
            print(key, value)
        elif key == "N":
            continue
        else:
            print(key, value.shape, value.dtype)

Hope this helps out!

[KyriaAnnwyn]

@hukkelas Great! Thank you for your quick answer!

[hukkelas]

My pleasure :)