We’ve all become used to deep learning’s success in image classification. Greater Swiss Mountain dog or Bernese mountain dog? Red panda or giant panda? No problem. However, in real life it’s not enough to name the single most salient object on a picture. Like it or not, one of the most compelling examples is autonomous driving: We don’t want the algorithm to recognize just that car in front of us, but also the pedestrian about to cross the street. And, just detecting the pedestrian is not sufficient. The exact location of objects matters.
The term object detection is commonly used to refer to the task of naming and localizing multiple objects in an image frame. Object detection is difficult; we’ll build up to it in a loose series of posts, focusing on concepts instead of aiming for ultimate performance. Today, we’ll start with a few straightforward building blocks: Classification, both single and multiple; localization; and combining both classification and localization of a single object.
Dataset
We’ll be using images and annotations from the Pascal VOC dataset which can be downloaded from this mirror. Specifically, we’ll use data from the 2007 challenge and the same JSON annotation file as used in the fast.ai course.
Quick download/organization instructions, shamelessly taken from a helpful post on the fast.ai wiki, are as follows:
# mkdir data && cd data
# curl -OL http://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
# curl -OL https://storage.googleapis.com/coco-dataset/external/PASCAL_VOC.zip
# tar -xf VOCtrainval_06-Nov-2007.tar
# unzip PASCAL_VOC.zip
# mv PASCAL_VOC/*.json .
# rmdir PASCAL_VOC
# tar -xvf VOCtrainval_06-Nov-2007.tarIn words, we take the images and the annotation file from different places:
- http://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar provides us with the images, and after unzipping all we care about is the
JPEGImagesfolder. - From https://storage.googleapis.com/coco-dataset/external/PASCAL_VOC.zip all we will be needing is the annotation file,
pascal_train2007.json.
Whether you’re executing the listed commands or arranging files manually, you should eventually end up with directories/files analogous to these:
img_dir <- "data/VOCdevkit/VOC2007/JPEGImages"
annot_file <- "data/pascal_train2007.json"
Now we need to extract some information from that json file.
Preprocessing
Let’s quickly make sure we have all required libraries loaded.
Annotations contain information about three types of things we’re interested in.
annotations <- fromJSON(file = annot_file)
str(annotations, max.level = 1)
List of 4
$ images :List of 2501
$ type : chr "instances"
$ annotations:List of 7844
$ categories :List of 20First, characteristics of the image itself (height and width) and where it’s stored. Not surprisingly, here it’s one entry per image.
Then, object class ids and bounding box coordinates. There may be multiple of these per image. In Pascal VOC, there are 20 object classes, from ubiquitous vehicles (car, aeroplane) over indispensable animals (cat, sheep) to more rare (in popular datasets) types like potted plant or tv monitor.
classes <- c(
"aeroplane",
"bicycle",
"bird",
"boat",
"bottle",
"bus",
"car",
"cat",
"chair",
"cow",
"diningtable",
"dog",
"horse",
"motorbike",
"person",
"pottedplant",
"sheep",
"sofa",
"train",
"tvmonitor"
)
boxinfo <- annotations$annotations %>% {
tibble(
image_id = map_dbl(., "image_id"),
category_id = map_dbl(., "category_id"),
bbox = map(., "bbox")
)
}
The bounding boxes are now stored in a list column and need to be unpacked.
For the bounding boxes, the annotation file provides x_left and y_top coordinates, as well as width and height. We will mostly be working with corner coordinates, so we create the missing x_right and y_bottom.
As usual in image processing, the y axis starts from the top.
Finally, we still need to match class ids to class names.
So, putting it all together:
imageinfo <- imageinfo %>%
inner_join(boxinfo, by = c("id" = "image_id")) %>%
inner_join(catinfo, by = c("category_id" = "id"))
Note that here still, we have several entries per image, each annotated object occupying its own row.
There’s one step that will bitterly hurt our localization performance if we later forget it, so let’s do it now already: We need to scale all bounding box coordinates according to the actual image size we’ll use when we pass it to our network.
target_height <- 224
target_width <- 224
imageinfo <- imageinfo %>% mutate(
x_left_scaled = (x_left / image_width * target_width) %>% round(),
x_right_scaled = (x_right / image_width * target_width) %>% round(),
y_top_scaled = (y_top / image_height * target_height) %>% round(),
y_bottom_scaled = (y_bottom / image_height * target_height) %>% round(),
bbox_width_scaled = (bbox_width / image_width * target_width) %>% round(),
bbox_height_scaled = (bbox_height / image_height * target_height) %>% round()
)
Let’s take a glance at our data. Picking one of the early entries and displaying the original image together with the object annotation yields
img_data <- imageinfo[4,]
img <- image_read(file.path(img_dir, img_data$file_name))
img <- image_draw(img)
rect(
img_data$x_left,
img_data$y_bottom,
img_data$x_right,
img_data$y_top,
border = "white",
lwd = 2
)
text(
img_data$x_left,
img_data$y_top,
img_data$name,
offset = 1,
pos = 2,
cex = 1.5,
col = "white"
)
dev.off()
Now as indicated above, in this post we’ll mostly address handling a single object in an image. This means we have to decide, per image, which object to single out.
A reasonable strategy seems to be choosing the object with the largest ground truth bounding box.
After this operation, we only have 2501 images to work with - not many at all! For classification, we could simply use data augmentation as provided by Keras, but to work with localization we’d have to spin our own augmentation algorithm. We’ll leave this to a later occasion and for now, focus on the basics.
Finally after train-test split
train_indices <- sample(1:n_samples, 0.8 * n_samples)
train_data <- imageinfo_maxbb[train_indices,]
validation_data <- imageinfo_maxbb[-train_indices,]
our training set consists of 2000 images with one annotation each. We’re ready to start training, and we’ll start gently, with single-object classification.
Single-object classification
In all cases, we will use XCeption as a basic feature extractor. Having been trained on ImageNet, we don’t expect much fine tuning to be necessary to adapt to Pascal VOC, so we leave XCeption’s weights untouched
and put just a few custom layers on top.
model <- keras_model_sequential() %>%
feature_extractor %>%
layer_batch_normalization() %>%
layer_dropout(rate = 0.25) %>%
layer_dense(units = 512, activation = "relu") %>%
layer_batch_normalization() %>%
layer_dropout(rate = 0.5) %>%
layer_dense(units = 20, activation = "softmax")
model %>% compile(
optimizer = "adam",
loss = "sparse_categorical_crossentropy",
metrics = list("accuracy")
)
How should we pass our data to Keras? We could simple use Keras’ image_data_generator, but given we will need custom generators soon, we’ll build a simple one ourselves. This one delivers images as well as the corresponding targets in a stream. Note how the targets are not one-hot-encoded, but integers - using sparse_categorical_crossentropy as a loss function enables this convenience.
batch_size <- 10
load_and_preprocess_image <- function(image_name, target_height, target_width) {
img_array <- image_load(
file.path(img_dir, image_name),
target_size = c(target_height, target_width)
) %>%
image_to_array() %>%
xception_preprocess_input()
dim(img_array) <- c(1, dim(img_array))
img_array
}
classification_generator <-
function(data,
target_height,
target_width,
shuffle,
batch_size) {
i <- 1
function() {
if (shuffle) {
indices <- sample(1:nrow(data), size = batch_size)
} else {
if (i + batch_size >= nrow(data))
i <<- 1
indices <- c(i:min(i + batch_size - 1, nrow(data)))
i <<- i + length(indices)
}
x <-
array(0, dim = c(length(indices), target_height, target_width, 3))
y <- array(0, dim = c(length(indices), 1))
for (j in 1:length(indices)) {
x[j, , , ] <-
load_and_preprocess_image(data[[indices[j], "file_name"]],
target_height, target_width)
y[j, ] <-
data[[indices[j], "category_id"]] - 1
}
x <- x / 255
list(x, y)
}
}
train_gen <- classification_generator(
train_data,
target_height = target_height,
target_width = target_width,
shuffle = TRUE,
batch_size = batch_size
)
valid_gen <- classification_generator(
validation_data,
target_height = target_height,
target_width = target_width,
shuffle = FALSE,
batch_size = batch_size
)
Now how does training go?
model %>% fit_generator(
train_gen,
epochs = 20,
steps_per_epoch = nrow(train_data) / batch_size,
validation_data = valid_gen,
validation_steps = nrow(validation_data) / batch_size,
callbacks = list(
callback_model_checkpoint(
file.path("class_only", "weights.{epoch:02d}-{val_loss:.2f}.hdf5")
),
callback_early_stopping(patience = 2)
)
)
For us, after 8 epochs, accuracies on the train resp. validation sets were at 0.68 and 0.74, respectively. Not too bad given given we’re trying to differentiate between 20 classes here.
Now let’s quickly think what we’d change if we were to classify multiple objects in one image. Changes mostly concern preprocessing steps.
Multiple object classification
This time, we multi-hot-encode our data. For every image (as represented by its filename), here we have a vector of length 20 where 0 indicates absence, 1 means presence of the respective object class:
image_cats <- imageinfo %>%
select(category_id) %>%
mutate(category_id = category_id - 1) %>%
pull() %>%
to_categorical(num_classes = 20)
image_cats <- data.frame(image_cats) %>%
add_column(file_name = imageinfo$file_name, .before = TRUE)
image_cats <- image_cats %>%
group_by(file_name) %>%
summarise_all(.funs = funs(max))
n_samples <- nrow(image_cats)
train_indices <- sample(1:n_samples, 0.8 * n_samples)
train_data <- image_cats[train_indices,]
validation_data <- image_cats[-train_indices,]
Correspondingly, we modify the generator to return a target of dimensions batch_size * 20, instead of batch_size * 1.
classification_generator <-
function(data,
target_height,
target_width,
shuffle,
batch_size) {
i <- 1
function() {
if (shuffle) {
indices <- sample(1:nrow(data), size = batch_size)
} else {
if (i + batch_size >= nrow(data))
i <<- 1
indices <- c(i:min(i + batch_size - 1, nrow(data)))
i <<- i + length(indices)
}
x <-
array(0, dim = c(length(indices), target_height, target_width, 3))
y <- array(0, dim = c(length(indices), 20))
for (j in 1:length(indices)) {
x[j, , , ] <-
load_and_preprocess_image(data[[indices[j], "file_name"]],
target_height, target_width)
y[j, ] <-
data[indices[j], 2:21] %>% as.matrix()
}
x <- x / 255
list(x, y)
}
}
train_gen <- classification_generator(
train_data,
target_height = target_height,
target_width = target_width,
shuffle = TRUE,
batch_size = batch_size
)
valid_gen <- classification_generator(
validation_data,
target_height = target_height,
target_width = target_width,
shuffle = FALSE,
batch_size = batch_size
)
Now, the most interesting change is to the model - even though it’s a change to two lines only. Were we to use categorical_crossentropy now (the non-sparse variant of the above), combined with a softmax activation, we would effectively tell the model to pick just one, namely, the most probable object.
Instead, we want to decide: For each object class, is it present in the image or not? Thus, instead of softmax we use sigmoid, paired with binary_crossentropy, to obtain an independent verdict on every class.
feature_extractor <-
application_xception(
include_top = FALSE,
input_shape = c(224, 224, 3),
pooling = "avg"
)
feature_extractor %>% freeze_weights()
model <- keras_model_sequential() %>%
feature_extractor %>%
layer_batch_normalization() %>%
layer_dropout(rate = 0.25) %>%
layer_dense(units = 512, activation = "relu") %>%
layer_batch_normalization() %>%
layer_dropout(rate = 0.5) %>%
layer_dense(units = 20, activation = "sigmoid")
model %>% compile(optimizer = "adam",
loss = "binary_crossentropy",
metrics = list("accuracy"))
And finally, again, we fit the model:
model %>% fit_generator(
train_gen,
epochs = 20,
steps_per_epoch = nrow(train_data) / batch_size,
validation_data = valid_gen,
validation_steps = nrow(validation_data) / batch_size,
callbacks = list(
callback_model_checkpoint(
file.path("multiclass", "weights.{epoch:02d}-{val_loss:.2f}.hdf5")
),
callback_early_stopping(patience = 2)
)
)
This time, (binary) accuracy surpasses 0.95 after one epoch already, on both the train and validation sets. Not surprisingly, accuracy is significantly higher here than when we had to single out one of 20 classes (and that, with other confounding objects present in most cases!).
Now, chances are that if you’ve done any deep learning before, you’ve done image classification in some form, perhaps even in the multiple-object variant. To build up in the direction of object detection, it is time we add a new ingredient: localization.
Single-object localization
From here on, we’re back to dealing with a single object per image. So the question now is, how do we learn bounding boxes? If you’ve never heard of this, the answer will sound unbelievably simple (naive even): We formulate this as a regression problem and aim to predict the actual coordinates. To set realistic expectations - we surely shouldn’t expect ultimate precision here. But in a way it’s amazing it does even work at all.
What does this mean, formulate as a regression problem? Concretely, it means we’ll have a dense output layer with 4 units, each corresponding to a corner coordinate.
So let’s start with the model this time. Again, we use Xception, but there’s an important difference here: Whereas before, we said pooling = "avg" to obtain an output tensor of dimensions batch_size * number of filters, here we don’t do any averaging or flattening out of the spatial grid. This is because it’s exactly the spatial information we’re interested in!
For Xception, the output resolution will be 7x7. So a priori, we shouldn’t expect high precision on objects much smaller than about 32x32 pixels (assuming the standard input size of 224x224).
Now we append our custom regression module.
We will train with one of the loss functions common in regression tasks, mean absolute error. But in tasks like object detection or segmentation, we’re also interested in a more tangible quantity: How much do estimate and ground truth overlap?
Overlap is usually measured as Intersection over Union, or Jaccard distance. Intersection over Union is exactly what it says, a ratio between space shared by the objects and space occupied when we take them together.
To assess the model’s progress, we can easily code this as a custom metric:
metric_iou <- function(y_true, y_pred) {
# order is [x_left, y_top, x_right, y_bottom]
intersection_xmin <- k_maximum(y_true[ ,1], y_pred[ ,1])
intersection_ymin <- k_maximum(y_true[ ,2], y_pred[ ,2])
intersection_xmax <- k_minimum(y_true[ ,3], y_pred[ ,3])
intersection_ymax <- k_minimum(y_true[ ,4], y_pred[ ,4])
area_intersection <- (intersection_xmax - intersection_xmin) *
(intersection_ymax - intersection_ymin)
area_y <- (y_true[ ,3] - y_true[ ,1]) * (y_true[ ,4] - y_true[ ,2])
area_yhat <- (y_pred[ ,3] - y_pred[ ,1]) * (y_pred[ ,4] - y_pred[ ,2])
area_union <- area_y + area_yhat - area_intersection
iou <- area_intersection/area_union
k_mean(iou)
}
Model compilation then goes like
Now modify the generator to return bounding box coordinates as targets…
localization_generator <-
function(data,
target_height,
target_width,
shuffle,
batch_size) {
i <- 1
function() {
if (shuffle) {
indices <- sample(1:nrow(data), size = batch_size)
} else {
if (i + batch_size >= nrow(data))
i <<- 1
indices <- c(i:min(i + batch_size - 1, nrow(data)))
i <<- i + length(indices)
}
x <-
array(0, dim = c(length(indices), target_height, target_width, 3))
y <- array(0, dim = c(length(indices), 4))
for (j in 1:length(indices)) {
x[j, , , ] <-
load_and_preprocess_image(data[[indices[j], "file_name"]],
target_height, target_width)
y[j, ] <-
data[indices[j], c("x_left_scaled",
"y_top_scaled",
"x_right_scaled",
"y_bottom_scaled")] %>% as.matrix()
}
x <- x / 255
list(x, y)
}
}
train_gen <- localization_generator(
train_data,
target_height = target_height,
target_width = target_width,
shuffle = TRUE,
batch_size = batch_size
)
valid_gen <- localization_generator(
validation_data,
target_height = target_height,
target_width = target_width,
shuffle = FALSE,
batch_size = batch_size
)
… and we’re ready to go!
model %>% fit_generator(
train_gen,
epochs = 20,
steps_per_epoch = nrow(train_data) / batch_size,
validation_data = valid_gen,
validation_steps = nrow(validation_data) / batch_size,
callbacks = list(
callback_model_checkpoint(
file.path("loc_only", "weights.{epoch:02d}-{val_loss:.2f}.hdf5")
),
callback_early_stopping(patience = 2)
)
)
After 8 epochs, IOU on both training and test sets is around 0.35. This number doesn’t look too good. To learn more about how training went, we need to see some predictions. Here’s a convenience function that displays an image, the ground truth box of the most salient object (as defined above), and if given, class and bounding box predictions.
plot_image_with_boxes <- function(file_name,
object_class,
box,
scaled = FALSE,
class_pred = NULL,
box_pred = NULL) {
img <- image_read(file.path(img_dir, file_name))
if(scaled) img <- image_resize(img, geometry = "224x224!")
img <- image_draw(img)
x_left <- box[1]
y_bottom <- box[2]
x_right <- box[3]
y_top <- box[4]
rect(
x_left,
y_bottom,
x_right,
y_top,
border = "cyan",
lwd = 2.5
)
text(
x_left,
y_top,
object_class,
offset = 1,
pos = 2,
cex = 1.5,
col = "cyan"
)
if (!is.null(box_pred))
rect(box_pred[1],
box_pred[2],
box_pred[3],
box_pred[4],
border = "yellow",
lwd = 2.5)
if (!is.null(class_pred))
text(
box_pred[1],
box_pred[2],
class_pred,
offset = 0,
pos = 4,
cex = 1.5,
col = "yellow")
dev.off()
img %>% image_write(paste0("preds_", file_name))
plot(img)
}
First, let’s see predictions on sample images from the training set.
train_1_8 <- train_data[1:8, c("file_name",
"name",
"x_left_scaled",
"y_top_scaled",
"x_right_scaled",
"y_bottom_scaled")]
for (i in 1:8) {
preds <-
model %>% predict(
load_and_preprocess_image(train_1_8[i, "file_name"],
target_height, target_width),
batch_size = 1
)
plot_image_with_boxes(train_1_8$file_name[i],
train_1_8$name[i],
train_1_8[i, 3:6] %>% as.matrix(),
scaled = TRUE,
box_pred = preds)
}
As you’d guess from looking, the cyan-colored boxes are the ground truth ones. Now looking at the predictions explains a lot about the mediocre IOU values! Let’s take the very first sample image - we wanted the model to focus on the sofa, but it picked the table, which is also a category in the dataset (although in the form of dining table). Similar with the image on the right of the first row - we wanted to it to pick just the dog but it included the person, too (by far the most frequently seen category in the dataset). So we actually made the task a lot more difficult than had we stayed with e.g., ImageNet where normally a single object is salient.
Now check predictions on the validation set.
Again, we get a similar impression: The model did learn something, but the task is ill defined. Look at the third image in row 2: Isn’t it pretty consequent the model picks all people instead of singling out some special guy?
If single-object localization is that straightforward, how technically involved can it be to output a class label at the same time? As long as we stay with a single object, the answer indeed is: not much.
Let’s finish up today with a constrained combination of classification and localization: detection of a single object.
Single-object detection
Combining regression and classification into one means we’ll want to have two outputs in our model. We’ll thus use the functional API this time. Otherwise, there isn’t much new here: We start with an XCeption output of spatial resolution 7x7, append some custom processing and return two outputs, one for bounding box regression and one for classification.
feature_extractor <- application_xception(
include_top = FALSE,
input_shape = c(224, 224, 3)
)
input <- feature_extractor$input
common <- feature_extractor$output %>%
layer_flatten(name = "flatten") %>%
layer_activation_relu() %>%
layer_dropout(rate = 0.25) %>%
layer_dense(units = 512, activation = "relu") %>%
layer_batch_normalization() %>%
layer_dropout(rate = 0.5)
regression_output <-
layer_dense(common, units = 4, name = "regression_output")
class_output <- layer_dense(
common,
units = 20,
activation = "softmax",
name = "class_output"
)
model <- keras_model(
inputs = input,
outputs = list(regression_output, class_output)
)
When defining the losses (mean absolute error and categorical crossentropy, just as in the respective single tasks of regression and classification), we could weight them so they end up on approximately a common scale. In fact that didn’t make much of a difference so we show the respective code in commented form.
model %>% freeze_weights(to = "flatten")
model %>% compile(
optimizer = "adam",
loss = list("mae", "sparse_categorical_crossentropy"),
#loss_weights = list(
# regression_output = 0.05,
# class_output = 0.95),
metrics = list(
regression_output = custom_metric("iou", metric_iou),
class_output = "accuracy"
)
)
Just like model outputs and losses are both lists, the data generator has to return the ground truth samples in a list. Fitting the model then goes as usual.
loc_class_generator <-
function(data,
target_height,
target_width,
shuffle,
batch_size) {
i <- 1
function() {
if (shuffle) {
indices <- sample(1:nrow(data), size = batch_size)
} else {
if (i + batch_size >= nrow(data))
i <<- 1
indices <- c(i:min(i + batch_size - 1, nrow(data)))
i <<- i + length(indices)
}
x <-
array(0, dim = c(length(indices), target_height, target_width, 3))
y1 <- array(0, dim = c(length(indices), 4))
y2 <- array(0, dim = c(length(indices), 1))
for (j in 1:length(indices)) {
x[j, , , ] <-
load_and_preprocess_image(data[[indices[j], "file_name"]],
target_height, target_width)
y1[j, ] <-
data[indices[j], c("x_left", "y_top", "x_right", "y_bottom")]
%>% as.matrix()
y2[j, ] <-
data[[indices[j], "category_id"]] - 1
}
x <- x / 255
list(x, list(y1, y2))
}
}
train_gen <- loc_class_generator(
train_data,
target_height = target_height,
target_width = target_width,
shuffle = TRUE,
batch_size = batch_size
)
valid_gen <- loc_class_generator(
validation_data,
target_height = target_height,
target_width = target_width,
shuffle = FALSE,
batch_size = batch_size
)
model %>% fit_generator(
train_gen,
epochs = 20,
steps_per_epoch = nrow(train_data) / batch_size,
validation_data = valid_gen,
validation_steps = nrow(validation_data) / batch_size,
callbacks = list(
callback_model_checkpoint(
file.path("loc_class", "weights.{epoch:02d}-{val_loss:.2f}.hdf5")
),
callback_early_stopping(patience = 2)
)
)What about model predictions? A priori we might expect the bounding boxes to look better than in the regression-only model, as a significant part of the model is shared between classification and localization. Intuitively, I should be able to more precisely indicate the boundaries of something if I have an idea what that something is.
Unfortunately, that didn’t quite happen. The model has become very biased to detecting a person everywhere, which might be advantageous (thinking safety) in an autonomous driving application but isn’t quite what we’d hoped for here.
Just to double-check this really has to do with class imbalance, here are the actual frequencies:
imageinfo %>% group_by(name)
%>% summarise(cnt = n())
%>% arrange(desc(cnt))# A tibble: 20 x 2
name cnt
<chr> <int>
1 person 2705
2 car 826
3 chair 726
4 bottle 338
5 pottedplant 305
6 bird 294
7 dog 271
8 sofa 218
9 boat 208
10 horse 207
11 bicycle 202
12 motorbike 193
13 cat 191
14 sheep 191
15 tvmonitor 191
16 cow 185
17 train 158
18 aeroplane 156
19 diningtable 148
20 bus 131To get better performance, we’d need to find a successful way to deal with this. However, handling class imbalance in deep learning is a topic of its own, and here we want to build up in the direction of objection detection. So we’ll make a cut here and in an upcoming post, think about how we can classify and localize multiple objects in an image.
Conclusion
We have seen that single-object classification and localization are conceptually straightforward. The big question now is, are these approaches extensible to multiple objects? Or will new ideas have to come in? We’ll follow up on this giving a short overview of approaches and then, singling in on one of those and implementing it.