search

zymk9 / yolov5_anime

Yet another anime face detector based on yolov5.
GNU General Public License v3.0
76 stars 5 forks source link

dataset for "bust" object #1

Open aperveyev opened 3 years ago

aperveyev commented 3 years ago

Inspired by EXCELLENT results of your face detector (weight model) yolov5s_anime.pt may I kindly ask you to check usability and quality of YOLO-compatible dataset for another class bust = pair of breasts + cleavage

dataset made (verified) using LabelImg classes.txt with one class "bust" 14475 JPGs named "imageboard-id.jpg" 14472 paired TXTs

images (1280 px height) taken from BOORU CHARS DATASET 2021 volume 2020, aspect ratios 1x2 and 3x4 processed with NudeNet used only subset (14.5k out of 77k) of images where "face + related pair of breasts" detected initial boundbox built over breasts area also using face bottom (and sometimes left-right) borders and then manually rechecked (not all but a lot of cases)

(MS One Drive, 3.7 GB) https://1drv.ms/u/s!Akw-XgJ0nd1-g4d2pUrWenBwBIfWVw?e=8fOR06

At the moment I'm not qualified enough to make models by myself (in Windows environment) but I'm working over it. Thanks in advance. Alex

aperveyev commented 3 years ago

My best model at the moment for 4 classes : 0 - head (face -> height + 60% upside, width + 2*25% bothsides - manually adjusted) 1 - bust (from collarbone center to pair of covered breasts included) 2 - bell (from belly button to thigh gap at full hips width, covered by something) 3 - boob [bust with no bra - often confuse with bust] https://1drv.ms/u/s!Akw-XgJ0nd1-g5BhER2P2q-nB1SbCg?e=u5yEVh produce good triplets 0 + 1/3 + 2 for great share of artbook pages

zymk9 commented 3 years ago

Hi Alex. Sorry that I'm not sure whether I can do any better in training and verification as I don't think I have a more "qualified" or suitable setting at hand.

Most of the work (except for annotation) for this face detection model was done using Google Colaboratory. You might find that a more friendly environment if you haven't tried for it.