make-dataset.py

# 1.unzip file
# 2. change *srcdir* 
# 3. run this script
# do training

import numpy as np
from pathlib import Path
import os

outdir = "datasets/data/"

# setwhich one
# srcdir = "corners"
# srcdir = "keypoints"

if srcdir=="keypoints":
    _KEYS = "start_kp center end_kp tip".split()
    DIM = 640
elif srcdir=="corners":
    _KEYS = "topLeft topRight bottomRight bottomLeft".split()
    DIM = 1280

F = os.listdir(srcdir) 

import random
random.seed(4)
random.shuffle(F)
F[0]

# 80% training
n = int(0.8*len(F))
n
train,val = F[:n], F[n:]

import shutil

for d in "images labels".split():
    for dd in "train val test".split():
        if not os.path.exists(f"{outdir}/{d}/{dd}"):
            os.makedirs(f"{outdir}/{d}/{dd}")

for trgd,F in [(f"{outdir}/images/train", train), (f"{outdir}/images/val", val), (f"{outdir}/images/test", val[:100])]:
    for srcf in F:
        shutil.move(os.path.join(srcdir, srcf), os.path.join(trgd, srcf))




def filename2keypoints(fn:str):
    kp = {}
    i = 0
    for sfn in fn.split("_"):
        if "-" in sfn:
            a,b = sfn.split("-")
            kp[_KEYS[i]] = (int(a),int(b))
            i+=1
    return kp


# <class-index> <x> <y> <width> <height> <px1> <py1> <px2> <py2> ... <pxn> <pyn>
def formatLabel(cx,cy, w,h, kp):
    pxs =""
    for k in _KEYS:
        x,y = kp[k]
        pxs += f"{x/DIM} {y/DIM} "
    return f"0 {cx/DIM} {cy/DIM} {w/DIM} {h/DIM} " +pxs.strip()

if srcdir=="keypoints":
    def toCenterCoordinates(kp,dim = DIM):
        # trying to determine bounding box via cv isnt reliable so we assume, dial face will always appears nearly same dim. and determine box from center point
        w, h = 560,560
        x,y = kp["center"]
        wh = w/2
        hh = h/2

        cx =  x-wh
        cy = y - hh # left top corner

        cx,cy, x,y
        _w,_h = 0,0
        # though this is not really correct, we will adjust dim. if bounding box is outside viewport
        if cx<0:
            _w = cx
        elif x+wh > dim:
            _w = (x+wh) - dim

        if cy < 0:
            _h = cy
        elif (y+hh)>dim:
            _h = (y+hh) - dim

        return x,y, w-_w,h-_h
else:
    def toCenterCoordinates(kp,dim = DIM):
        # bounding corners
        c = np.array(list(kp.values()))
        margin = 10
        topLeft,bottomRight = c.min(axis=0)-margin, c.max(axis=0)+margin
        topLeft,bottomRight =  np.clip(topLeft, 0, dim),np.clip(bottomRight, 0, dim)
        cx,cy = 0.5*(bottomRight + topLeft)
        w,h = bottomRight-topLeft
        return cx,cy, w,h


# make annotation file
for d,F in [(f"{outdir}/images/train", train), (f"{outdir}/images/val", val), (f"{outdir}/images/test", val[:100])]:
    for f in F:
        _,ext = os.path.splitext(f)
        antfn = os.path.join(outdir, d.replace("images/", "labels/"), _+".txt")
        kp = filename2keypoints(f)
        cx,cy,w,h = toCenterCoordinates(kp)
        assert cx>0
        assert cy>0
        L = formatLabel(cx,cy,w,h, kp)
        with open(antfn, "w") as fn:
            fn.write(L+"\n")