In [1]:
import os
import sys
import itertools
import math
import logging
import json
import re
import random
from collections import OrderedDict
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import matplotlib.lines as lines
from matplotlib.patches import Polygon
# Root directory of the project
ROOT_DIR = os.path.abspath("../../")
# Import Mask RCNN
sys.path.append(ROOT_DIR) # To find local version of the library
from mrcnn import utils
from mrcnn import visualize
from mrcnn.visualize import display_images
import mrcnn.model as modellib
from mrcnn.model import log
from samples.balloon import balloon
%matplotlib inline
In [2]:
config = balloon.BalloonConfig()
BALLOON_DIR = os.path.join(ROOT_DIR, "datasets/balloon")
In [3]:
# Load dataset
# Get the dataset from the releases page
# https://github.com/matterport/Mask_RCNN/releases
dataset = balloon.BalloonDataset()
dataset.load_balloon(BALLOON_DIR, "train")
# Must call before using the dataset
dataset.prepare()
print("Image Count: {}".format(len(dataset.image_ids)))
print("Class Count: {}".format(dataset.num_classes))
for i, info in enumerate(dataset.class_info):
print("{:3}. {:50}".format(i, info['name']))
In [4]:
# Load and display random samples
image_ids = np.random.choice(dataset.image_ids, 4)
for image_id in image_ids:
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
visualize.display_top_masks(image, mask, class_ids, dataset.class_names)
Rather than using bounding box coordinates provided by the source datasets, we compute the bounding boxes from masks instead. This allows us to handle bounding boxes consistently regardless of the source dataset, and it also makes it easier to resize, rotate, or crop images because we simply generate the bounding boxes from the updates masks rather than computing bounding box transformation for each type of image transformation.
In [5]:
# Load random image and mask.
image_id = random.choice(dataset.image_ids)
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
# Compute Bounding box
bbox = utils.extract_bboxes(mask)
# Display image and additional stats
print("image_id ", image_id, dataset.image_reference(image_id))
log("image", image)
log("mask", mask)
log("class_ids", class_ids)
log("bbox", bbox)
# Display image and instances
visualize.display_instances(image, bbox, mask, class_ids, dataset.class_names)
In [6]:
# Load random image and mask.
image_id = np.random.choice(dataset.image_ids, 1)[0]
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
original_shape = image.shape
# Resize
image, window, scale, padding, _ = utils.resize_image(
image,
min_dim=config.IMAGE_MIN_DIM,
max_dim=config.IMAGE_MAX_DIM,
mode=config.IMAGE_RESIZE_MODE)
mask = utils.resize_mask(mask, scale, padding)
# Compute Bounding box
bbox = utils.extract_bboxes(mask)
# Display image and additional stats
print("image_id: ", image_id, dataset.image_reference(image_id))
print("Original shape: ", original_shape)
log("image", image)
log("mask", mask)
log("class_ids", class_ids)
log("bbox", bbox)
# Display image and instances
visualize.display_instances(image, bbox, mask, class_ids, dataset.class_names)
Instance binary masks can get large when training with high resolution images. For example, if training with 1024x1024 image then the mask of a single instance requires 1MB of memory (Numpy uses bytes for boolean values). If an image has 100 instances then that's 100MB for the masks alone.
To improve training speed, we optimize masks by:
To visualize the effect of mask resizing, and to verify the code correctness, we visualize some examples.
In [7]:
image_id = np.random.choice(dataset.image_ids, 1)[0]
image, image_meta, class_ids, bbox, mask = modellib.load_image_gt(
dataset, config, image_id, use_mini_mask=False)
log("image", image)
log("image_meta", image_meta)
log("class_ids", class_ids)
log("bbox", bbox)
log("mask", mask)
display_images([image]+[mask[:,:,i] for i in range(min(mask.shape[-1], 7))])
In [8]:
visualize.display_instances(image, bbox, mask, class_ids, dataset.class_names)
In [9]:
# Add augmentation and mask resizing.
image, image_meta, class_ids, bbox, mask = modellib.load_image_gt(
dataset, config, image_id, augment=True, use_mini_mask=True)
log("mask", mask)
display_images([image]+[mask[:,:,i] for i in range(min(mask.shape[-1], 7))])
In [10]:
mask = utils.expand_mask(bbox, mask, image.shape)
visualize.display_instances(image, bbox, mask, class_ids, dataset.class_names)
The order of anchors is important. Use the same order in training and prediction phases. And it must match the order of the convolution execution.
For an FPN network, the anchors must be ordered in a way that makes it easy to match anchors to the output of the convolution layers that predict anchor scores and shifts.
Anchor Stride: In the FPN architecture, feature maps at the first few layers are high resolution. For example, if the input image is 1024x1024 then the feature meap of the first layer is 256x256, which generates about 200K anchors (2562563). These anchors are 32x32 pixels and their stride relative to image pixels is 4 pixels, so there is a lot of overlap. We can reduce the load significantly if we generate anchors for every other cell in the feature map. A stride of 2 will cut the number of anchors by 4, for example.
In this implementation we use an anchor stride of 2, which is different from the paper.
In [11]:
# Generate Anchors
backbone_shapes = modellib.compute_backbone_shapes(config, config.IMAGE_SHAPE)
anchors = utils.generate_pyramid_anchors(config.RPN_ANCHOR_SCALES,
config.RPN_ANCHOR_RATIOS,
backbone_shapes,
config.BACKBONE_STRIDES,
config.RPN_ANCHOR_STRIDE)
# Print summary of anchors
num_levels = len(backbone_shapes)
anchors_per_cell = len(config.RPN_ANCHOR_RATIOS)
print("Count: ", anchors.shape[0])
print("Scales: ", config.RPN_ANCHOR_SCALES)
print("ratios: ", config.RPN_ANCHOR_RATIOS)
print("Anchors per Cell: ", anchors_per_cell)
print("Levels: ", num_levels)
anchors_per_level = []
for l in range(num_levels):
num_cells = backbone_shapes[l][0] * backbone_shapes[l][1]
anchors_per_level.append(anchors_per_cell * num_cells // config.RPN_ANCHOR_STRIDE**2)
print("Anchors in Level {}: {}".format(l, anchors_per_level[l]))
Visualize anchors of one cell at the center of the feature map of a specific level.
In [12]:
## Visualize anchors of one cell at the center of the feature map of a specific level
# Load and draw random image
image_id = np.random.choice(dataset.image_ids, 1)[0]
image, image_meta, _, _, _ = modellib.load_image_gt(dataset, config, image_id)
fig, ax = plt.subplots(1, figsize=(10, 10))
ax.imshow(image)
levels = len(backbone_shapes)
for level in range(levels):
colors = visualize.random_colors(levels)
# Compute the index of the anchors at the center of the image
level_start = sum(anchors_per_level[:level]) # sum of anchors of previous levels
level_anchors = anchors[level_start:level_start+anchors_per_level[level]]
print("Level {}. Anchors: {:6} Feature map Shape: {}".format(level, level_anchors.shape[0],
backbone_shapes[level]))
center_cell = backbone_shapes[level] // 2
center_cell_index = (center_cell[0] * backbone_shapes[level][1] + center_cell[1])
level_center = center_cell_index * anchors_per_cell
center_anchor = anchors_per_cell * (
(center_cell[0] * backbone_shapes[level][1] / config.RPN_ANCHOR_STRIDE**2) \
+ center_cell[1] / config.RPN_ANCHOR_STRIDE)
level_center = int(center_anchor)
# Draw anchors. Brightness show the order in the array, dark to bright.
for i, rect in enumerate(level_anchors[level_center:level_center+anchors_per_cell]):
y1, x1, y2, x2 = rect
p = patches.Rectangle((x1, y1), x2-x1, y2-y1, linewidth=2, facecolor='none',
edgecolor=(i+1)*np.array(colors[level]) / anchors_per_cell)
ax.add_patch(p)
In [13]:
# Create data generator
random_rois = 2000
g = modellib.data_generator(
dataset, config, shuffle=True, random_rois=random_rois,
batch_size=4,
detection_targets=True)
In [14]:
# Uncomment to run the generator through a lot of images
# to catch rare errors
# for i in range(1000):
# print(i)
# _, _ = next(g)
In [15]:
# Get Next Image
if random_rois:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_class_ids, gt_boxes, gt_masks, rpn_rois, rois], \
[mrcnn_class_ids, mrcnn_bbox, mrcnn_mask] = next(g)
log("rois", rois)
log("mrcnn_class_ids", mrcnn_class_ids)
log("mrcnn_bbox", mrcnn_bbox)
log("mrcnn_mask", mrcnn_mask)
else:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_boxes, gt_masks], _ = next(g)
log("gt_class_ids", gt_class_ids)
log("gt_boxes", gt_boxes)
log("gt_masks", gt_masks)
log("rpn_match", rpn_match, )
log("rpn_bbox", rpn_bbox)
image_id = modellib.parse_image_meta(image_meta)["image_id"][0]
print("image_id: ", image_id, dataset.image_reference(image_id))
# Remove the last dim in mrcnn_class_ids. It's only added
# to satisfy Keras restriction on target shape.
mrcnn_class_ids = mrcnn_class_ids[:,:,0]
In [16]:
b = 0
# Restore original image (reverse normalization)
sample_image = modellib.unmold_image(normalized_images[b], config)
# Compute anchor shifts.
indices = np.where(rpn_match[b] == 1)[0]
refined_anchors = utils.apply_box_deltas(anchors[indices], rpn_bbox[b, :len(indices)] * config.RPN_BBOX_STD_DEV)
log("anchors", anchors)
log("refined_anchors", refined_anchors)
# Get list of positive anchors
positive_anchor_ids = np.where(rpn_match[b] == 1)[0]
print("Positive anchors: {}".format(len(positive_anchor_ids)))
negative_anchor_ids = np.where(rpn_match[b] == -1)[0]
print("Negative anchors: {}".format(len(negative_anchor_ids)))
neutral_anchor_ids = np.where(rpn_match[b] == 0)[0]
print("Neutral anchors: {}".format(len(neutral_anchor_ids)))
# ROI breakdown by class
for c, n in zip(dataset.class_names, np.bincount(mrcnn_class_ids[b].flatten())):
if n:
print("{:23}: {}".format(c[:20], n))
# Show positive anchors
fig, ax = plt.subplots(1, figsize=(16, 16))
visualize.draw_boxes(sample_image, boxes=anchors[positive_anchor_ids],
refined_boxes=refined_anchors, ax=ax)
In [17]:
# Show negative anchors
visualize.draw_boxes(sample_image, boxes=anchors[negative_anchor_ids])
In [18]:
# Show neutral anchors. They don't contribute to training.
visualize.draw_boxes(sample_image, boxes=anchors[np.random.choice(neutral_anchor_ids, 100)])
In [19]:
if random_rois:
# Class aware bboxes
bbox_specific = mrcnn_bbox[b, np.arange(mrcnn_bbox.shape[1]), mrcnn_class_ids[b], :]
# Refined ROIs
refined_rois = utils.apply_box_deltas(rois[b].astype(np.float32), bbox_specific[:,:4] * config.BBOX_STD_DEV)
# Class aware masks
mask_specific = mrcnn_mask[b, np.arange(mrcnn_mask.shape[1]), :, :, mrcnn_class_ids[b]]
visualize.draw_rois(sample_image, rois[b], refined_rois, mask_specific, mrcnn_class_ids[b], dataset.class_names)
# Any repeated ROIs?
rows = np.ascontiguousarray(rois[b]).view(np.dtype((np.void, rois.dtype.itemsize * rois.shape[-1])))
_, idx = np.unique(rows, return_index=True)
print("Unique ROIs: {} out of {}".format(len(idx), rois.shape[1]))
In [20]:
if random_rois:
# Dispalay ROIs and corresponding masks and bounding boxes
ids = random.sample(range(rois.shape[1]), 8)
images = []
titles = []
for i in ids:
image = visualize.draw_box(sample_image.copy(), rois[b,i,:4].astype(np.int32), [255, 0, 0])
image = visualize.draw_box(image, refined_rois[i].astype(np.int64), [0, 255, 0])
images.append(image)
titles.append("ROI {}".format(i))
images.append(mask_specific[i] * 255)
titles.append(dataset.class_names[mrcnn_class_ids[b,i]][:20])
display_images(images, titles, cols=4, cmap="Blues", interpolation="none")
In [21]:
# Check ratio of positive ROIs in a set of images.
if random_rois:
limit = 10
temp_g = modellib.data_generator(
dataset, config, shuffle=True, random_rois=10000,
batch_size=1, detection_targets=True)
total = 0
for i in range(limit):
_, [ids, _, _] = next(temp_g)
positive_rois = np.sum(ids[0] > 0)
total += positive_rois
print("{:5} {:5.2f}".format(positive_rois, positive_rois/ids.shape[1]))
print("Average percent: {:.2f}".format(total/(limit*ids.shape[1])))
In [ ]: