4.6 - Object detection

!wget -nc --no-cache -O init.py -q https://raw.githubusercontent.com/rramosp/2021.deeplearning/main/content/init.py
import init; init.init(force_download=False);

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import pandas as pd
%matplotlib inline
%load_ext tensorboard

from sklearn.datasets import *
from local.lib import mlutils
from IPython.display import Image
from skimage import io
tf.__version__

!wget -nc https://storage.googleapis.com/openimages/v5/class-descriptions-boxable.csv
c = pd.read_csv("class-descriptions-boxable.csv", names=["code", "description"], index_col="code")
c.head()

img = io.imread("local/imgs/0003bb040a62c86f.jpg")
plt.figure(figsize=(12,10))
plt.imshow(img)

boxes = pd.read_csv("local/data/openimages_boxes_0003bb040a62c86f.csv")
boxes

pd.Series([c.loc[i].description for i in boxes.LabelName]).value_counts()

from matplotlib.patches import Rectangle
i = np.random.randint(len(boxes))
plt.figure(figsize=(12,10));
ax = plt.subplot(111)
plt.imshow(img)
h,w = img.shape[:2]
for i in range(len(boxes)):
    k = boxes.iloc[i]
    label = c.loc[k.LabelName].values[0]
    ax.add_patch(Rectangle((k.XMin*w,k.YMin*h),(k.XMax-k.XMin)*w,(k.YMax-k.YMin)*h, linewidth=2,edgecolor='r',facecolor='none'))
    plt.text(k.XMin*w, k.YMin*h-10, label, fontsize=12, color="red")

patches = [img[190:300, 150:270],
           img[200:300, 600:700],
           img[400:500, 400:500],
           img[200:300, 300:400],
           img[220:290, 325:360],
           img[10:180, 330:670]]

plt.figure(figsize=(20,3))
for i,pimg in enumerate(patches):
    plt.subplot(1,len(patches),i+1); plt.imshow(pimg)

from tensorflow.keras.applications import inception_v3
if not "model" in locals():
    model = inception_v3.InceptionV3(weights='imagenet', include_top=True)

def plot_img_with_histogram(img):
    plt.figure(figsize=(13,4))
    plt.subplot(121)
    plt.imshow(img, vmin=np.min(img), vmax=np.max(img))
    plt.subplot(122)
    plt.hist(img.flatten(), bins=30);

pimg = patches[2]

plot_img_with_histogram(pimg)

from skimage.transform import resize
rimg = resize(pimg, output_shape=(299,299,3))
plot_img_with_histogram(rimg)

pred = model.predict(rimg.reshape(-1,*rimg.shape))
pred.shape

k = pd.DataFrame(inception_v3.decode_predictions(pred, top=100)[0], columns=["code", "label", "preds"])
k = k.sort_values(by="preds", ascending=False)

plt.figure(figsize=(15,3))
n = 40
plt.bar(range(n), k[:n].preds.values)
plt.xticks(range(n), k[:n].label.values, rotation="vertical");
plt.title("classes with highest prediction probability")
plt.grid();

print ('Predicted:')
k = inception_v3.decode_predictions(pred, top=10)[0]
for i in k:
    print("%10s %20s %.6f"%i)

import tensorflow_hub as hub

classnames = pd.read_csv('https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt', names=["label"])

if not 'm' in locals():
  m = hub.KerasLayer("https://tfhub.dev/google/imagenet/inception_resnet_v2/classification/4")

[w.shape for w in m.weights]

preds = m(rimg.reshape(-1,*rimg.shape).astype(np.float32)).numpy()[0]
preds = np.exp(preds)/np.sum(np.exp(preds))
np.sum(preds)

names = classnames.copy()
names["preds"] = preds
names = names.sort_values(by="preds", ascending=False)
names.head()

plt.figure(figsize=(15,3))
n = 40
plt.bar(range(n), names[:n].preds.values)
plt.xticks(range(n), names[:n].label.values, rotation="vertical");
plt.title("classes with highest prediction probability")
plt.grid();

Image("local/imgs/anchor_boxes.png", width=300)

fig = plt.figure(figsize=(8,8))
ax = plt.subplot(111)
plt.imshow(img)
n = 9
for i in range(n):
    plt.axvline(img.shape[1]//n*i, color="black")
    plt.axhline(img.shape[0]//n*i, color="black")

k = boxes.iloc[0]
label = c.loc[k.LabelName].values[0]
ax.add_patch(Rectangle((k.XMin*w,k.YMin*h),(k.XMax-k.XMin)*w,(k.YMax-k.YMin)*h,
                       linewidth=4,edgecolor='r',facecolor='none'))
plt.text(k.XMin*w, k.YMin*h-10, label, fontsize=12, color="red")
plt.scatter(k.XMin*w+(k.XMax-k.XMin)*w*.5,k.YMin*h+(k.YMax-k.YMin)*h*.5, color="red", s=100)

Image("local/imgs/yolo_predictions.png")