Reasons for better results about person-reid-triplet-loss-baseline

[huanghoujing]

I think the point is

• longer training time
• smaller margin, 0.5 -> 0.3

Other differences may also matter

• I do not add additional fully connected layers after the ResNet backbone
• I do not apply random cropping at training time

Thanks for your watching. If you have any interesting finding in your experiments, feel free to share it in the issues.

[liangbh6]

Well, I found that for batch-hard triplet loss

1. better result will be obtained without fc added;
2. smaller batch size will get inferior results, eg, batch_size=72 is better than batch_size=32, but the runtime for smaller batch is shorter.
   Your results are attractive. Thanks for your kind reply~

[huanghoujing]

Thanks for sharing! Someone may find these useful.