did you check the original polygon for BYLS-107 and plot it with the original image you will see the mask is not aligned it's almost 90 degree rotation . base don my understanding the polygon and mask should be overlapped

I did for almost all and was confused at first. But when I visualized it on the map it actually made sense. All those polygons fit in the Bare land perfectly. I then realized most of the polygons were from different site plans.

There might be one or two I overlooked but it is not really significant to drop your score to low.

Have you tried flipping the y-axis again before visualizing?

it's not all about flipping.

I did inference on all training set images and there are those samples you can never get right even though my model is ok.

i totally agree with you because i did the same thing i infere on training data and the mask perfectly fit the polygons but evaluating the iou between predicted masks and ground truth gives you low iou

Also this second plot shows how real polygon masks are imperfect if you cross check with actual images you will note all the different irregularites present.

thank you @Joseph_gitau for you info and confirmation for the same point i mentioned may i ask you if you don't mind plot BYLS-107 specially because this one has totally wrong mask just confirmation thank you again

I see, this is really different.

Well if Amy says everything is correct, we just have to take her words for it. I mean like people will not be in 97and 96 polygon scores

They would be even with wrong polygons

I think the issue stems from using the image pixels to get the gt polygons, note that the polygons provided are in geo coords, so there will be some small disrepancies - as it is not a 1 to 1 mapping(pixel poly vs geo coords poly)- this could explain the 0.95 iou, but never 1. but if you use bearings and distances to get the gt polygons, most of the polygons are correct, though there is still a number of images with wrong polgons.

This are all my assumptions and made up theories 😂

This is what I have

Exactly, this was my point @Brainiac.

This is where domain knowledge could have worked wonders

Hi @Brainiac i spent all the day working on bearing and distances to get the gt masks i use affine transfoermation and georefernce points but same issue i know this is the tight time but can you explain a methodology or piepline to use bearing and distance to get exact gt polygons for images with correct shapes , thanks

Hi @Mohamed_abdelrazik — totally get where you’re coming from. I went down the same rabbit hole with cadastral maps; it took me a while (papers, articles, YouTube deep dives) to really grasp how bearings/distances tie back to ground truth. It’s a new domain, so one day isn’t realistic—don’t be hard on yourself.

Here’s the pipeline I use for bearings + distances:

1. Locate the EN point in the image Pick the parcel’s reference point “EN” in the image as the origin for the walk.
2. Walk clockwise from EN with bearings & distances Read each boundary leg in order (clockwise). For each leg, use its bearing and distance to get the next vertex. Keep stepping around the parcel until you return to EN.
3. Get the polygon coords from the provided function Feed the ordered vertex list (starting at EN, clockwise, closed) into the provided bearings_to_en_coords function to generate the mask/polygon with the correct shape.

Example image:

import math

from typing import List, Tuple

import numpy as np

import matplotlib.pyplot as plt

Bearing = Tuple[float, float, float]  # (deg, minutes, distance)

EN = Tuple[float, float]  # (E, N)

def bearings_to_en_coords(

    bearings: List[Bearing],

    start_en: EN,

    use_back: bool = True,

    shift: int = 0,

    reverse: bool = False,

    bowditch: bool = True,

    close: bool = True,

) -> List[EN]:

    if reverse:

        bearings = list(reversed(bearings))

    if shift:

        k = shift % len(bearings)

        bearings = bearings[k:] + bearings[:k]

    de, dn, Ls = [], [], []

    for d, m, L in bearings:

        az = (d + m / 60.0) % 360.0

        if use_back:

            az = (az + 180.0) % 360.0

        r = math.radians(az)

        de.append(L * math.sin(r))

        dn.append(L * math.cos(r))

        Ls.append(L)

    de = np.asarray(de, float)

    dn = np.asarray(dn, float)

    Ls = np.asarray(Ls, float)

    if bowditch and Ls.sum() > 0:

        ce, cn = de.sum(), dn.sum()

        w = Ls / Ls.sum()

        de -= ce * w

        dn -= cn * w

    x = np.r_[0.0, de.cumsum()]

    y = np.r_[0.0, dn.cumsum()]

    E0, N0 = start_en

    pts = [(E0 + float(xi), N0 + float(yi)) for xi, yi in zip(x, y)]

    if close and pts[-1] != pts[0]:

        pts.append(pts[0])

    return pts

def plot_coords(coords: List[EN], title: str = "Converted EN Polygon"):

    if not coords:

        raise ValueError("Empty coords.")

    xs = [p[0] for p in coords]

    ys = [p[1] for p in coords]

    plt.figure(figsize=(6, 6))

    plt.plot(xs, ys, "-o")

    ax = plt.gca()

    ax.set_aspect("equal", adjustable="box")

    ax.grid(True, ls="--", alpha=0.3)

    ax.set_xlabel("Easting")

    ax.set_ylabel("Northing")

    ax.set_title(title)

    plt.tight_layout()

    plt.show()

USE_BACK = True

bearings = [

    (45, 27, 6.10),

    (10, 0, 4.69),

    (6, 43, 3.96),

    (96, 42, 24.56),

    (135, 37, 20.67),

    (225, 42, 28.32),

    (315, 29, 30.51),]

start_en = (40836.55, 74893.59)

coords_en = bearings_to_en_coords(

    bearings, start_en, use_back=USE_BACK, bowditch=True, close=True

)

plot_coords(

    coords_en, title=f"Converted EN ({'Back' if USE_BACK else 'Forward'} bearings)"

)

@Brainiac Thank you Big man🙇

@Brainiac THank you so much for your time and help bro