search

Huangdebo / Confluence

A Robust Non-IoU Alternative to Non-Maxima Suppression in Object Detection
49 stars 10 forks source link

Differences from paper #7

Open danielcrane opened 2 years ago

danielcrane commented 2 years ago

Sorry for asking in English, but I have a couple of questions regarding your implementation of Confluence.

1.

May I ask where the following line comes from in the paper?

https://github.com/Huangdebo/Confluence/blob/8347895175179e1acc222826594eb576047de4b5/confluence.py#L167

In equation (6) of the paper, they weight the confidence score like so:

image

where P is the Manhattan distance, and s is the confidence, however in your case it looks more like you're doing:

P_w = P / s

I suppose the outcome will likely be the same, but I was just wondering if this method was mentioned in the paper, as I can't see it?

2.

In section 3.5 of the paper, after selecting the optimal bounding box, they then decay the confidences of all bounding boxes within close proximity (p_thres) of this optimal bounding box.

image

As far as I can see, in your implementation this stage was skipped, and instead your just delete all bounding boxes within p_thres of the optimal bounding box:

https://github.com/Huangdebo/Confluence/blob/8347895175179e1acc222826594eb576047de4b5/confluence.py#L188-L190

May I ask why you decided to do this instead? Did you run some tests and this method performed better?

Thanks so much in advance!

Huangdebo commented 2 years ago

Sorry for just replying you now.

My implementation is based on an earlier version of the paper int ARXIV. In this version, the equation (6) was

image

and the pseudocode was image

By the way，the line 12 in the pseudocode is Proximity * si which is not math the equation (6). Maybe in the version of officially published paper , these two points had been optimized .