Starship Detection
Solution for submission 152146
A detailed solution for submission 152146 submitted for challenge Starship Detection
ksnxr
Starter Code for Starship Detection
What we are going to Learn¶
- Basic Image Preprocessing with OpenCV
- Fixing Contours on an image and getting bounding boxes.
- Testing and Submitting the Results to the Challenge.
Note : Create a copy of the notebook and use the copy for submission. Go to File > Save a Copy in Drive to create a new copy
Setting up Environment¶
Downloading Dataset¶
So we will first need to download the python library by AIcrowd that will allow us to download the dataset by just inputting the API key.
In [1]:
!pip install aicrowd-cli
%load_ext aicrowd.magic
In [2]:
%aicrowd login --api-key c18a6c4b759fa2df29481af274df6942
In [3]:
# Downloading the Dataset
!rm -rf data
!mkdir data
%aicrowd ds dl -c starship-detection -o data
In [4]:
!unzip data/data.zip -d /content/images >> /dev/null
In [5]:
!pip uninstall scikit-image -y
!pip install scikit-image
Importing Libraries¶
In [6]:
# Image Reading & Preprocessing
import skimage.io
import matplotlib.pyplot as plt
import numpy as np
# Misc.
import pandas as pd
from tqdm.notebook import tqdm
import os
from natsort import natsorted
from skimage.color import rgb2gray
from skimage.filters import sobel
from skimage.draw import rectangle
from skimage.measure import find_contours
Image Preprocessing¶
In this section we are going to learn some opencv functions which can help us detecting the starship body!
In [116]:
# Reading a Sample Image
img = skimage.io.imread('images/6625.jpg')
plt.imshow(img)
Out[116]:
In [117]:
gray = rgb2gray(img)
plt.imshow(gray)
Out[117]:
In [123]:
result = sobel(gray)
plt.imshow(result)
Out[123]:
In [124]:
contours = find_contours(result, 0.0001)
def get_bounding_rect(contour):
x_min = np.min(contour[:, 0])
x_max = np.max(contour[:, 0])
y_min = np.min(contour[:, 1])
y_max = np.max(contour[:, 1])
return round(x_min), round(y_min), round(x_max), round(y_max)
def draw_rect(img, points, color=[255, 0, 0]):
# Make sure the coordinates are expressed as integers
rr, cc = rectangle(start=points[:2], end=points[2:])
img[rr, cc] = color
return img
x_min = 512
y_min = 512
x_max = 0
y_max = 0
draw_img = img.copy()
for contour in contours:
points = get_bounding_rect(contour)
x_min = min(x_min, points[0])
y_min = min(y_min, points[1])
x_max = max(x_max, points[2])
y_max = max(y_max, points[3])
print(x_min, y_min, x_max, y_max)
draw_img = draw_rect(draw_img, [x_min, y_min, x_max, y_max])
plt.imshow(draw_img)
Out[124]:
Countours¶
Contours are lines joining along the bounding of a intensity or color in an image. In the canny image or the original image, we see that the starship has much different color as compared to the sky.
So as you can see, fnding contours did a pretty great job in finfing the starship body. However there are some mistakes with the image in such as the more right side of the starship body in left unchecked.
Submitting Results 📄¶
In [53]:
# Getting all the testing images
test_imgs = natsorted(os.listdir("images"))
print(len(test_imgs))
In [113]:
# Function to generate the bounding boxes
def gen_bounding_boxes(img):
# Converting the image to numpy array
gray = rgb2gray(img)
result = sobel(gray)
contours = find_contours(result, 0.05)
def get_bounding_rect(contour):
x_min = np.min(contour[:, 0])
x_max = np.max(contour[:, 0])
y_min = np.min(contour[:, 1])
y_max = np.max(contour[:, 1])
return round(x_min), round(y_min), round(x_max), round(y_max)
def draw_rect(img, points, color=[255, 0, 0]):
# Make sure the coordinates are expressed as integers
rr, cc = rectangle(start=points[:2], end=points[2:])
img[rr, cc] = color
return img
x_min = 512
y_min = 512
x_max = 0
y_max = 0
draw_img = img.copy()
for contour in contours:
points = get_bounding_rect(contour)
x_min = min(x_min, points[0])
y_min = min(y_min, points[1])
x_max = max(x_max, points[2])
y_max = max(y_max, points[3])
return y_min, x_min, y_max - y_min, x_max - x_min
In [114]:
bboxes = []
image_ids = []
# Ground through each test image
for img_name in tqdm(test_imgs):
# Reading the test image
img = skimage.io.imread(os.path.join("images", img_name))
# Generating the bounding boxes
x,y,w,h = gen_bounding_boxes(img)
# Adding the boundong boxes and image id
bboxes.append([x,y,w,h])
image_ids.append(int(img_name.split(".")[0]))
In [127]:
fill = True
xs = []
ys = []
ws = []
hs = []
missing_idxes = []
for idx, bbox in enumerate(bboxes):
x, y, w, h = bbox
if x == 512 or x == 0 or y == 512 or y == 0 or w <=0 or h <= 0:
print(idx)
missing_idxes.append(idx)
else:
xs.append(x)
ys.append(y)
ws.append(w)
hs.append(h)
if fill:
x_mean = sum(xs) / len(xs)
y_mean = sum(ys) / len(ys)
w_mean = sum(ws) / len(ws)
h_mean = sum(hs) / len(hs)
for idx in missing_idxes:
bboxes[idx] = [x_mean, y_mean, w_mean, h_mean]
In [ ]:
from PIL import Image, ImageDraw
def visualize(index):
image_id = image_ids[index]
bbox = bboxes[index]
x, y, w, h = bbox
draw_img = Image.open(f"images/{image_id}.jpg")
draw = ImageDraw.Draw(draw_img)
draw.rectangle([x, y, x+w, y+h], outline ="red")
return draw_img
In [96]:
visualize(182)
Out[96]:
In [79]:
visualize(2)
Out[79]:
In [80]:
visualize(100)
Out[80]:
In [ ]:
def merge_boxes(boxes):
In [81]:
# Adding the image id and bounding boxes to a dataframe
df = pd.DataFrame({"ImageID":image_ids, "bbox":bboxes})
df = df.sort_values("ImageID").reset_index(drop=True)
df
Out[81]:
In [86]:
!rm -rf assets
!mkdir assets
df.to_csv(os.path.join("assets", "submission.csv"), index=False)
Note : Please make sure that there should be filename submission.csv in assets folder before submitting it
Uploading the Results¶
In [91]:
In [ ]:
Content
Comments
You must login before you can post a comment.