In detectron2, how could I plot only bounding on images without label (class name) in object detection?

Question

I am learning detectron2 and practicing it using pokemonster data. So, I follow the detectron2 tutorial (here : https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5)

And also a detector is developed and below image can be detected.

But, in this image, red circle is class name. I want to plot only bounding box(green color) with the exception of red circle (class name). Now below is my code for visualization. What code should be modified? Thank you.

import glob
for imageName in sorted(glob.glob(os.path.join(test_path, '*.jpg'))):
  im = cv2.imread(imageName)
  outputs = predictor(im)
  v = Visualizer(im[:, :, ::-1],
                metadata=train_metadata,
                scale=0.8
                 )
  out = v.draw_instance_predictions(outputs["instances"].to("cpu"))
  boxes = v._convert_boxes(outputs["instances"].pred_boxes.to('cpu'))
  for box in boxes:
    box = (round(box[0]), round(box[1]), round(box[2]) - round(box[0]), round(box[3] - box[1]))
    out = v.draw_text(f"{box[2:4]}", (box[0], box[1]))
  cv2_imshow(out.get_image()[:, :, ::-1])

Alexander Riedel · Accepted Answer · 2020-11-18T09:44:02.580

4

The class Visualizer contains all the methods inlcuding draw_box(). You want to draw all the boxes to your Visualizer object and finally return a VisImage

v = Visualizer(
        im[:, :, ::-1], 
        metadata=train_metadata, 
        scale=0.8,
        )
for box in outputs["instances"].pred_boxes.to('cpu'):
    v.draw_box(box)
    v.draw_text(str(box[:2].numpy()), tuple(box[:2].numpy()))
v = v.get_output()
img =  v.get_image()[:, :, ::-1]
cv2_imshow(img)

edited Nov 18 '20 at 09:44

answered Nov 18 '20 at 09:13

Alexander Riedel

1,329
1
7
14

Thank you for commenting. But your code only show bounding box. I want to plot bounding box coordinate like above with bounding box. Could you please help me more? – Seunghyeon Nov 18 '20 at 09:26
Yes sure, i updated my answer, you can draw whatever you want... – Alexander Riedel Nov 18 '20 at 09:43

score 0 · Answer 2 · answered Apr 02 '21 at 13:18

There is a function to draw binary mask, but there is no function to draw mask on the colored image. How to draw mask on the original input color image.

output = predictor(input_image)        
V = Visualizer(input_image[:, :, ::-1], 
MetadataCatalog.get(cfg.DATASETS.TRAIN[0]), scale = 1.2)
mask=output["instances"].get("pred_masks")
mask=mask.to("cpu")
V.draw_instance_predictions(output["instances"].to("cpu"))
for m in mask:
    V.draw_mask(m)

V = V.get_output()
cv2.imwrite("frame%d.jpg" % count, V.get_image()[:, :, ::-1])

Thank you

score 0 · Answer 3 · answered Jun 14 '22 at 09:55

Here is a working version of boxes and masks:

v = Visualizer(im[:, :, ::-1],
               metadata=train_metadata, 
               scale=0.8,
)
# Boxes
for box in outputs["instances"].pred_boxes.to('cpu'):
    v.draw_box(box)    
# Masks
for mask in outputs["instances"].pred_masks.to('cpu'):
    v.draw_soft_mask(mask)
v = v.get_output()
img =  v.get_image()[:, :, ::-1]
cv2_imshow(img)

In detectron2, how could I plot only bounding on images without label (class name) in object detection?

3 Answers3