search

WongKinYiu / yolor

implementation of paper - You Only Learn One Representation: Unified Network for Multiple Tasks (https://arxiv.org/abs/2105.04206)
GNU General Public License v3.0
1.99k stars 518 forks source link

What makes the difference of AP between the paper and this repository? #205

Open yukitsuji opened 2 years ago

yukitsuji commented 2 years ago

AP values reported in this repository are much higher than the ones reported in the paper. What makes the difference of AP between the paper and this repository?

Model Test Size APtest AP50test AP75test batch1 throughput batch32 inference
YOLOR-CSP 640 52.8% 71.2% 57.6% 106 fps 3.2 ms
YOLOR-CSP-X 640 54.8% 73.1% 59.7% 87 fps 5.5 ms
YOLOR-P6 1280 55.7% 73.3% 61.0% 76 fps 8.3 ms
YOLOR-W6 1280 56.9% 74.4% 62.2% 66 fps 10.7 ms
YOLOR-E6 1280 57.6% 75.2% 63.0% 45 fps 17.1 ms
YOLOR-D6 1280 58.2% 75.8% 63.8% 34 fps 21.8 ms
WongKinYiu commented 2 years ago

We apply this in our new training scheme.

yukitsuji commented 2 years ago

Thanks. @WongKinYiu Would you have a plan to open your internal code to reproduce the new training scheme?

WongKinYiu commented 2 years ago

yes, we are preparing the code.

kevinvlowell commented 2 years ago

Hi @WongKinYiu ! Are there any updates regarding the readiness of this code?

Are the pre-trained weights for the APs reported in this repo available anywhere for download?