search

Went-Liang / UnSniffer

[CVPR 2023] Official Pytorch code for Unknown Sniffer for Object Detection: Don’t Turn a Blind Eye to Unknown Objects
https://arxiv.org/pdf/2303.13769v1.pdf
Apache License 2.0
67 stars 9 forks source link

None #13

Closed YH-2023 closed 12 months ago

YH-2023 commented 1 year ago

I have a question. Since the GOC branch can detect whether the input feature is an object and only includes known and unknown classes, what is the purpose of the classification branch in the second detector? The GOC module itself has the ability to detect known objects, so is the concatenation redundant? Can't we achieve both known and unknown object detection using just the regression and GOC branches? Is it also possible to directly infer from the results obtained by GOC?

Went-Liang commented 1 year ago

Usually, more accurate supervision information allows the model to learn more precise knowledge. GOC is designed mainly to address the training challenges posed by unknown objects. In contrast, the classification branch in the second detector benefits from richer supervision information, so we have retained it.

YH-2023 commented 1 year ago

@Went-Liang Thank you very much for your prompt reply I added an additional goc branch (corresponding to the first detector in the paper) based on the original classification and regression (corresponding to the second detector in the paper). Before adding the goc module, the known class It can reach 55, but after I added the goc module, the known mAP was only 47. I don’t understand why it affects the known mAP. I also retained the cls branch and regression branch. It stands to reason that the known mAP should remain unchanged. Which aspect do you think is causing the known accuracy to decrease? I have been thinking about this for a long time.

Went-Liang commented 1 year ago

I'm sorry, but I can't understand your description. Could you please draw the model structure for me?

Went-Liang commented 1 year ago

I'm guessing you are using GOC scores in the detection classification branch of known objects? This score is not accurate compared with the score of the classification branch, so in our paper, the GOC score is only used to filter objects in the unknown object detection branch.

YH-2023 commented 1 year ago

I'm guessing you are using GOC scores in the detection classification branch of known objects? This score is not accurate compared with the score of the classification branch, so in our paper, the GOC score is only used to filter objects in the unknown object detection branch.

In my model, I use GOC, cls, and reg to train loss together.Because the network I use is relatively simple, there are no complex post-processing operations during inference.