paper related problem

[megvii-model]

paper related problem

[chensnathan]

Hi, thanks for your questions.

In MiMo encoders, there are five levels of features, from P3 to P7. The receptive field in P6 and P7 is relatively larger than C5, making the receptive field problem not severe. However, in SiSo encoders, there is only one level feature (C5 or DC5). We have to enlarge the receptive field, compensating for the lack of P6 and P7.

While for small objects, it is true that C5 achieves an inferior performance in detecting small objects (Shown in Table1, YOLOF(19.1 mAP) vs. RetinaNet+(22.2 mAP)). However, it is hard to recover more detailed information from a high-level feature, thus we try to keep the detailed information in the C5 feature by adopting a shortcut.

We also show a feasible way to improve the detection performance on small objects in Table 9 in the paper. We use the dilated C5 feature to replace the C5 feature. The performance of small objects can be improved (from 19.1 mAP(YOLOF-C5-1x) to 22.3 mAP(YOLOF-DC5-1x)).

[chensnathan]

Yes, the performance drop is mainly caused by small objects.

The results are list below:
MiMo:
AP | AP50 | AP75 | APs | APm | APl |
35.9 | 55.8 | 38.5 | 19.9 | 39.6 | 47.9 |

SiMo:
AP | AP50 | AP75 | APs | APm | APl |
35.0 | 54.8 | 37.1 | 17.4 | 39.3 | 47.8 |

[yypurpose]

Thanks for your timely and detailed answer!

So it's bridging the gap between SiSo and SiMo, not SiSo and MiMo! My understanding was wrong. And dilated conversations is to make up for the missing of P6 & P7. I DO UNDERTAND THIS TIME.

And as you have said in the paper, SiMo is comparable with MiMo (35.0mAP vs. 35.9mAP), is the main drop caused by the small object or it's uniform? Could you provide the detailed results of the four models in Fig. 1 as I did not found them in the paper. I'm really curious about it.

THANKS A LOT!!

[yypurpose]

Thanks a lot!