iccv19 light weight face recognition challenge tips

[nttstar]

Challenge Information:

HomePage, please read carefully about the training and testing rules.

How To Start:

Please check https://github.com/deepinsight/insightface/tree/master/iccv19-challenge for detail.

Any question can be left here.

[borisgribkov]

HI @nttstar, my traditional question regarding the dataset, is it already 5-keypoints aligned? Sorry, have not downloaded it.

[nttstar]

@borisgribkov Yes already aligned, so no need to further process on it.

[zhouwei5113]

@nttstar what's the difference between ms1m-retinaface-t1.zip and ms1m-v2(emore)?

[nttstar]

@zhouwei5113 ms1m-retinaface has different clean list than ms1m-arcface and it has been re-aligned with our RetinaFace detector. You must use ms1m-retinaface to participate in this challenge, all other training-sets including ms1m-arcface are not allowed.

[leolordwz]

Hi, @nttstar, could you please share with us how the provided dataset generated from ms-1m aligned dataset. I tried to upscale images in ms-1m aligned dataset to 400x400, but the landmarks do not match. It would be great if you could share with us the preparation process.

[nttstar]

@leolordwz it's not allowed for this challenge.

[chenshen03]

How can I take part in the iccv19 light weight face recognition challenge?

[nttstar]

@chenshen03 Test datasets and test server will be ready soon, any question on training?

[jamessmith90]

@nttstar Can you provide more info about RetinaFace Detector ?

[LCorleone]

@nttstar Is the test set aligned by RetinaFace？ If not, will you release the RetinaFace Detector？ Many thanks！

[nttstar]

@LCorleone They're both aligned by RetinaFace, so no need to do anything on training or testing datasets.

[nttstar]

@BingranHu We can get higher accuracy on LFW, CFP_FP and AgeDB_30 by the help of ms1m-retinaface.

[xizi]

@nttstar Hi, where is the download links of test data?

[nttstar]

@xizi Not ready yet.

[GewelsJI]

Does the organizing association have some limitation on the number of participants? @nttstar

[nttstar]

@GewelsJI No

[grib0ed0v]

@nttstar When do you expect test server will be ready?

[nttstar]

@grib0ed0v Soon~

[sybyun]

I have a question about the size of parameters. Does the 20M parameters mean total (include FC layers) size or only backbone network's size?

[nttstar]

@sybyun From the input layer to embedding layer. By the way, it is 20MB size which means 5M parameters.

[chenshen03]

The test server could not log in.

[nttstar]

@chenshen03 what's the problem? I can login successfully.

[chenshen03]

@nttstar The account can login normally on http://trillionpairs.deepglint.com, but there was an error when it logged on to the test server http://39.104.128.76.

[nttstar]

@chenshen03 sign up a new account, database not shared with deepglint server.

[chenshen03]

@nttstar The database is not shared, but the system prompts "user name or email already exists" when registering. Do we need to use a different email address to register?

Also, when I register with a new email box, I get an "unknown error". When I try to register again, "user or email already exists" will appear.

[nttstar]

@chenshen03 Hi, what is your old email address?

[chenshen03]

@nttstar email: chenshen@stu.xmu.edu.cn username: Shen Chen. Thank you.

[nttstar]

@chenshen03 Can you try again to sign up a new account with chenshen@stu.xmu.edu.cn ? We have done a cleaning.

[chenshen03]

@nttstar Successfully registered, thank you!

[yeluyue]

@nttstar I couldn't find where to download test data's pair list. Can you tell me the download link of the pair list for making the binary feature matrix (ImageNum x FeatureDim in float32) to the test server?

[nttstar]

@yeluyue Pair list is not required.

[yeluyue]

@nttstar Then how to generate the binary feature matrix? Now I have downloaded the test data and the feature_tools from the trillion Pairs, and I think the images needing to make feature matrix exists the insightface.bin. Can you tell me whether my understanding is correct?

[nttstar]

@yeluyue Please check the tutorial here. I have provided the feature generation tools.

[yeluyue]

@nttstar Thank you ~

[tianxingyzxq]

Entity validate failed: user name is empty when i sumbit the feature bin. @nttstar

[nttstar]

@tianxingyzxq Try logout and login again?

[NothingL]

Is data augmentation(random resize and crop) permitted?

[nttstar]

@NothingL It is allowed. But you must always use pre-aligned test dataset so normally you will get a worse result by using training time random resize and crop.

[ckybit]

hi @nttstar ，it is too slow to upload the bin file, 7 hours passed and it only upload 75%. is the right ? how long do you take to upload the 1.8G bin file ? thank you very much.

[nttstar]

@ckybit less than 5 minutes. where are you?

[ckybit]

@ckybit less than 5 minutes. where are you? i'm in china and i use company's network. i will change the network to try again,thks for your quick reply.

[zhouwei5113]

@nttstar Is there any approach to upload test bin file on my training server directly, other than by browser?

[nttstar]

@zhouwei5113 Sorry no command line tool

[NothingL]

@NothingL It is allowed. But you must always use pre-aligned test dataset so normally you will get a worse result by using training time random resize and crop.

thanks for clarification, is that mean we cannot resize and crop test image, even the cropped size is also 112*112 ?(means using a smaller region of face)

[yeluyue]

Is there any differences between the feature matrices that need to be submitted for deepglint-light and deepglint-large, or should the feature matrices be the same for both of them? @nttstar

[YongdongTan]

How do I open the .z01 and .z02 files to get the image? I am using a ubuntu system.

[nttstar]

@NothingL Right

[nttstar]

@yeluyue The shape of submission feature should always be (N, feature_dim) N=num of images. feature_dim can be 512, 256, etc..

[nttstar]

@YongdongTan Please check the testing section of tutorial.

[olojuwin]

how to get Network y2 and Network r100fc?