Mask_rcnn: Demo with multiple images

[racinmat]

Hi, I really like the demo which makes it super easy to segment some image, but I fail to use it for multiple images at once.
I thought that modifying predicting in demo.ipynb

file_names = next(os.walk(IMAGE_DIR))[2]
image = skimage.io.imread(os.path.join(IMAGE_DIR, random.choice(file_names)))

# Run detection
results = model.detect([image], verbose=1)

# Visualize results
r = results[0]
visualize.display_instances(image, r['rois'], r['masks'], r['class_ids'], 
                            class_names, r['scores'])

to that:

file_names = next(os.walk(IMAGE_DIR))[2]
image0 = skimage.io.imread(os.path.join(IMAGE_DIR, random.choice(file_names)))
image1 = skimage.io.imread(os.path.join(IMAGE_DIR, random.choice(file_names)))
# image = skimage.io.imread(os.path.join(IMAGE_DIR, ''))

# plt.figure(figsize=(14, 14))
# plt.axis('off')
# plt.imshow(image.astype(np.uint8))
# plt.show()
# Run detection
results = model.detect([image0, image1], verbose=1)

# Visualize results
r = results[0]
visualize.display_instances(image0, r['rois'], r['masks'], r['class_ids'], 
                            class_names, r['scores'])
r = results[1]
visualize.display_instances(image1, r['rois'], r['masks'], r['class_ids'], 
                            class_names, r['scores'])

would result in predicting for two images, but instead, segmentation and bounding boxes from first image are used also for the second one:

[hrahoy]

I used plt.cla() to flush my canvas.

[racinmat]

Did not work for me, it is not problem of visualization, I tried to display only the second result ad second image, code is:

image0 = skimage.io.imread(os.path.join(IMAGE_DIR, random.choice(file_names)))
image1 = skimage.io.imread(os.path.join(IMAGE_DIR, random.choice(file_names)))

results = model.detect([image0, image1], verbose=1)

# Visualize results
# r = results[0]
# visualize.display_instances(image0, r['rois'], r['masks'], r['class_ids'], 
#                            class_names, r['scores'])
r = results[1]
visualize.display_instances(image1, r['rois'], r['masks'], r['class_ids'], 
                            class_names, r['scores'])

displaying of first image is commented out.

[waleedka]

This is odd. Did you change the batch size well?

[racinmat]

No, I haven't.
And it seemd that was the problem. Such stupid mistake.
After changing batch size to

class InferenceConfig(coco.CocoConfig):
    # Set batch size to 1 since we'll be running inference on
    # one image at a time. Batch size = GPU_COUNT * IMAGES_PER_GPU
    GPU_COUNT = 1
    IMAGES_PER_GPU = 2

it works perfectly. Thanks.

[waleedka]

Great!

By the way, get the latest version. I added an assert statement that catches this particular case and notifies you if the batch size is wrong.

[zqdeepbluesky]

@racinmat
hi,I have a question about demo.py,If I have a folder,and there are a lot of pictures under this folder, I want to test all of them, what should I do?can you show me how to do it? thanks so much.

[superabhijha2000]

how to load images without randomly . i want to load image as a specific image.

[racinmat]

If you have a lots of images, more than you can fit at once, then then you need to call detection multiple times, one for each batch.
You can probably split images into batches, which will make it faster, or just feed the whole directory into the network one image by one.

And when you don't want to load image randomly, then simply remove the random part of code a put specific image here.