search

WongKinYiu / yolov9

Implementation of paper - YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information
GNU General Public License v3.0
9.04k stars 1.44k forks source link

Question about 5.4.2 in paper and reversible branch #507

Open MinkiSong opened 5 months ago

MinkiSong commented 5 months ago

First, thank you for the excellent paper and code release. I have some questions as I would like to understand your paper and code more deeply.

Section 5.4.2 of the paper conducts an ablation study on PGI.

image

There are mentions of ICN, PFH, FPN, and PAN here. Could you provide references for ICN and PFH? There are no descriptions of G_backbone and G_neck in Table 4. Could you explain how they differ? In section 4.1, it states, "PGI mainly includes three components, namely (1) main branch, (2) auxiliary reversible branch, and (3) multi-level auxiliary information." Therefore, I understand GELAN + PGI as applying "auxiliary reversible branch" and "multi-level auxiliary information" to GELAN.

The performance in the repository's readme appears to be the same as the performance of GELAN + PGI in Table 5. Thus, I understand that all configs of yolov9 have "auxiliary reversible branch" and "multi-level auxiliary information" applied. image image

However, based on the conversation in issue #196, it seems that only yolov9-e has the "auxiliary reversible branch" applied.

Is the auxiliary reversible branch only applied to yolov9-e? If not, where does the "auxiliary reversible branch" apply in yolov9-S, M, and C?

yjtlab commented 3 months ago

same question

MinkiSong commented 2 months ago

@WongKinYiu Can you answer this question? Or can you tag someone who can answer?