Personal question

[nadongjin]

Hello
Happy New Year!
This is a personal question.
What architecture should we use to make facial recognition accuracy sota performance?
Based on MS1MV3 data, EfficientNetV2S or EffV2S, and AdamW's architecture are small in size, so the information speed is good, but the accuracy is slightly poor.
Personally, I would like to ask you what architecture is good to use to improve accuracy even if the speed is slightly slow.
Is r100 the best choice? ^^;
Thank you always for your kind reply.

[leondgarse]

I‘m mostly trying efficientnetv2 recently. Testing a strategy parameters on EffV2B0 then apply to EffV2S. As far as I tested, EffV2S is slightly better than r50 on MS1MV3:

• My EfficientNetV2S + MS1MV3 + SGD IJBB 95.6475, IJBC 96.8605.
• My se_r50 swish SD + MS1MV3 + SGD IJBB 95.6378, IJBC 96.8144.
• insightface model_zoo IResNet-50 + MS1MV3 IJBC 96.580.

r100 is another level on model size, so comparing with r50 here.
But generally talking which is better, IMO EffcientNets are good ones in most CV transfer learning missions, while ResNet family models have a good potential enhancing with other structures like se / bottleneck attention / halo attention or others. Like I have a ResNetV250 like model using anti alias down sample instead of MaxPooling and PReLU activation trained using MS1MV3 + SGD reached IJBC 96.8656, but it's still not a stable architecture.
Further, there are also other models for CV missions like ConvNeXt / CoAtNet and many others can be found in image-classification-on-imagenet.
Currently, I would just recommend EfficientNets till found some others better. Yes, r100 is still better than Effv2S, and my testing on EffV2M not as good as expect. Using MS1MV3 + SGD trained 50 epochs, Effv2M is IJBB 0.95667, IJBC 0.969781, just a slightly better than V2S.