I trained instance segmentation,but I don't have bbox about mask2former HOT 15 OPEN

Instance segmentation is not object detection and thus does not require bbox predictions. But you can always draw a bbox around mask, check issue: #10 (comment)

from mask2former.

@bowenc0221 Hi, I have bbox score.Thank you!
I want to ask a question. where should I set confidence-threshold when I demo my dataset?I set the confidence-threshold=0.5 in demo.py, but always detected 100 instances in one Image.It seems an invalid parameter?

from mask2former.

You can also do instances = instances[instances.scores > SCORE_THRESHOLD]

from mask2former.

@bowenc0221 Thanks you!!

from mask2former.

@FrankFang0813 Hi . I want to do a training with my own efficiency, but I have no idea how to do it. How should your data set be and can you share the training codes with me?

from mask2former.

@yasarniyazoglu Hi,I not sure what you mean.You can't train your custom dataset? Or you have same problem that just have mask AP but don't have bbox AP?

from mask2former.

@yasarniyazoglu Hi,I not sure what you mean.You can't train your custom dataset? Or you have same problem that just have mask AP but don't have bbox AP?

I can't train with private data. and I don't know how to do it either. Do you have training codes?

from mask2former.

@bowenc0221 Hi, I have bbox score.Thank you! I want to ask a question. where should I set confidence-threshold when I demo my dataset?I set the confidence-threshold=0.5 in demo.py, but always detected 100 instances in one Image.It seems an invalid parameter?

Hello, I met the same problem. Have you sovled it?

from mask2former.

hello which config did you use?

from mask2former.

@bowenc0221 Hi, I have bbox score.Thank you! I want to ask a question. where should I set confidence-threshold when I demo my dataset?I set the confidence-threshold=0.5 in demo.py, but always detected 100 instances in one Image.It seems an invalid parameter?

Its not obvious to me that confidence-threshold actually does something. I cannot find it being used anywhere in the code.

from mask2former.

If you want to set a score to filter the predictions, you can consider doing like below

from mask2former.

You can also do instances = instances[instances.scores > SCORE_THRESHOLD]

No matter how many instances there are, the size of the model prediction output is always 100.

from mask2former.

@SEUZTh
Hi.
Do you know of any method to make the model prediction output have less than 100 instances?

Cheers

from mask2former.

I have solved this issue. here is my solution:

`

     num_instances = mask_array.shape[0]
        scores = predictions["instances"].scores.cpu().numpy()
        for i in scores:
            if i < args.confidence_threshold: 
                num_instances -= 1
                num_instances = max(num_instances, 0)
        print('num_instances:', num_instances)

`

from mask2former.

@SEUZTh Hi. Do you know of any method to make the model prediction output have less than 100 instances?

Cheers

if "instances" in predictions:
    instances = predictions["instances"].to(self.cpu_device) # 类型: <class 'detectron2.structures.instances.Instances'>
    # 取得分大于阈值的实例
    instances_ = Instances(instances.image_size)
    flag = False
    for index in range(len(instances)):
        print(instances[index].scores)
        score = instances[index].scores[0]
        if score > 0.75: # 置信度设置
            if flag == False:
                instances_ = instances[index]
                flag = True
            else:
                instances_ = Instances.cat([instances_, instances[index]])
    vis_output = visualizer.draw_instance_predictions(predictions=instances_)

More information, you can refer to my blog: https://blog.csdn.net/qq_44324181/article/details/126242948

from mask2former.