Hide / Remove bounding Box at Inference

[hacadtra]

Hi,
Is there a way to not display bounding box and class name when inferencing model?
In my case, I will only count the total detection and show only the instance segmentation of the given image. I only use 1 class for this detection.

normal result

desired result

for the normal result I've used this code:

im = cv2.imread(d["file_name"])
outputs = predictor(I'm)
v = Visualizer(im[:, :, ::-1], metadata=microcontroller_metadata, scale=0.8)
v = v.draw_instance_predictions(outputs["instances"].to("cpu"))
plt.figure(figsize = (14, 10))
plt.imshow(cv2.cvtColor(v.get_image()[:, :, ::-1], cv2.COLOR_BGR2RGB))
plt.show()

hope you can give some suggestions

[VGrondin]

Yes you can, do you still need it?

[hacadtra]

yes, please enlighten me

[VGrondin]

You can use overlay_instances. Here is a code snippet:

# visualize
im = cv2.imread(d["file_name"])
# predictions
outputs_pred = predictor(im)   
v = Visualizer(img,
                metadata=metadata, 
                scale=1, 
                instance_mode =  ColorMode.BW     # remove color from image, better see segmentation mask  
)
dt_image = v.overlay_instances(
                boxes=None, 
                masks=outputs_pred["instances"].pred_masks.to("cpu"),  
)
        
cv2.imshow('predictions', dt_image[:, :, ::-1])
cv2.waitKey(0)

If your image colors are reversed, remove the [:, :, ::-1]

[sushmasuresh28]

@VGrondin this totally helped! Thank you so much :)

With no class name, the results are as follows-

After I manually register(as you suggested), I get exactly what I need-

Out of curiosity, I am trying to compute the area enclosed by each of the predicted polygons. I have been looking to access the coordinates of the drawn polygons but can't seem to find how they are stored.

By running print(outputs["instances"].pred_boxes) I get a list of tensors like this -

Do you know what this information would actually mean?

The documentation here - https://detectron2.readthedocs.io/en/latest/tutorials/models.html suggests that the "instance" object has"pred_masks" which should contain "a Tensor of shape (N, H, W), masks for each detected instance".

But when I run print(outputs["instances"].pred_masks) I get -
. Not sure why this is the case when clearly I am able to see the masks on the image.

I don't understand how to obtain the polygon information to calculate the area. If you know how the information can be accessed, could you please help me?

[VGrondin]

Glad it worked!

The mask format you printed is a binary image. I guess it the fastest method for models during inference time, otherwise it would have to encode the binary image into polygon/rle format which would slow down the process. Anyways binary images are easy, to handle:

# print number of true pixels (area) for each i predicted instances
for i in range(len(outputs["instances"].pred_masks)):
    print(torch.sum(outputs["instances"].pred_masks[i]))

[sushmasuresh28]

@VGrondin This totally makes sense! Thank you for clarifying :)

I just have one final question that I can't seem to find the answer to that anywhere(including the documentation here).

As shown above, when I run print(outputs["instances"].pred_boxes) I get a list of tensors like this -

Are the printed tensors representing the edge coordinates of each predicted bounding box? If yes, why are they float values/ decimals? If they are not the bounding box points, then what exactly are they?
If you happen to know what pred_boxes is actually representing and can share that with me, it will surely make my understanding of this concept better

[IAmOZRules]

hi
@VGrondin can you pls suggest how to get the coordinates of the polygon. As you suggested torch.sum gives the area but for my application, I need the X,Y coordinates of each of the polygon points. Thanks

[VGrondin]

# convert binary mask to polygon mask
has_mask = instances.has("pred_masks")
if has_mask:
     segmentations = []
     areas = []
     for mask in instances.pred_masks:
         polygons = Mask(mask.numpy()).polygons()
         segmentations.append(polygons.segmentation)
         
         area = Mask(mask.numpy()).area()
         areas.append(int(area))`