In detectron2 there are class IDs instead of class names

Question

I finished training model for instance segmentation in detectron2 when I test images in training files there is no problem class names(apple,banana,orange) are written on the image but I downloaded some fruit images from the internet and class names are not written on the photos. There are class ID's.

cfg.MODEL.WEIGHTS = os.path.join(cfg.OUTPUT_DIR, "model_final.pth")
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5 
cfg.DATASETS.TEST = ("fruit_test", )
predictor = DefaultPredictor(cfg)

image_path = "/content/detectron2_custom_dataset/testimages/test2.jpg"

def on_image(image_path,predictor):
    im = cv2.imread(image_path)
    outputs = predictor(im)
    v = Visualizer(im[:,:,::-1], metadata = {}, scale=0.5, instance_mode = ColorMode.SEGMENTATION)
    v = v.draw_instance_predictions(outputs["instances"].to("cpu"))
    plt.figure(figsize=(14,10))
    plt.imshow(v.get_image())
    plt.show()

on_image(image_path, predictor)

In conclusion, I want test my model with I uploaded now and I don't want to have class id on the images. I want class names like orange,banana,apple on it

Answer 1

After searching for an easy solution, I came with this, which is the easiest IMO.

Create a new Metadata class

Class Metadata:
    def get(self, _):
        return ['apple','banana','orange','etc'] #your class labels

then, provide Metadata in the Visualizer line

v = Visualizer(im[:, :, ::-1], Metadata, scale=0.5, instance_mode = ColorMode.SEGMENTATION)

Otherwise, you're required to register a model with (a dummy) data or even 1 image with annotations, then load its metadata. Which I find a bit irrelevant at this stage, if you need it for the inference code only.

Answer 2

You can get the labels by populating the metadata kwarg, which contains the mapping.

v = Visualizer(im[:, :, ::-1], MetadataCatalog.get(cfg.DATASETS.TEST[0]), scale=0.5, instance_mode = ColorMode.SEGMENTATION)