AVA v2.1 pre-trained baseline model

[Chen Sun]

Hi AVA Users,

I'm happy to announce the release of pre-trained baseline model trained on AVA v2.1.

The baseline model is an image-based Faster RCNN detector with ResNet-101 feature extractor. It was created using the Tensorflow Object Detection API.

Compared with other commonly used object detectors, the action classification loss function has been changed to per-class sigmoid loss to handle boxes with multiple labels. The model was trained on the training split of AVA v2.1 for 1.5M iterations, and achieves mean AP of 11.25% over 60 classes on the validation split of AVA v2.1.

The model checkpoint can be obtained here. The predictions of this model on the AVA v2.1 validation split, in the CSV format described above, can be downloaded here: ava_baseline_detections_val_v2.1.zip.

Chen Sun

On behalf of the whole AVA team

[João Antunes]

Hello, AVA team,

I just wanted to ask, what is the Actor detection performance of the baseline you provided now? Is it the reported 76% we find on table 5 of the AVA paper?

Thanks!

João

[kv...@uw.edu]

Hi AVA team,

Thank you for releasing the checkpoints and predictions in CSV format. However, I can not download the CSV file. 

Error messages: ava_baseline_detections_val_v2.1.zip’s server IP address could not be found.

Thanks,

Kevin

[Chen Sun]

Hi João,

Yes, it's 75.3% as shown in Figure 5 (on the latest version of our arXiv paper)

Chen

--
You received this message because you are subscribed to the Google Groups "AVA Dataset Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ava-dataset-us...@googlegroups.com.
To post to this group, send email to ava-data...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ava-dataset-users/514f387c-d46b-4cd0-862d-b84d08ca6b14%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Chen Sun]

Hi Kevin,

Can you please try again? I just tried this link and can download the zip file.

Chen

--

You received this message because you are subscribed to the Google Groups "AVA Dataset Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ava-dataset-us...@googlegroups.com.
To post to this group, send email to ava-data...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/ava-dataset-users/1914c13e-c38b-40ed-84e1-169e42dd1e2b%40googlegroups.com.

[Vijay Kumar]

Hi AVA team,

Is it possible to share the model checkpoint and config file instead of .pb file of your baseline faster RCNN Resnet101 model? We want to extract proposal features of this model, and it seems your .pb file is generated with old version of tensorflow and it doesn't output features. 

There is a solution suggested in the following post, where they suggest to convert old tensorflow models to newer versions. The newer version will directly output proposal features. However such conversion can happen only with .checkpoint and .config files.  

https://stackoverflow.com/questions/49170336/tf-object-detection-api-extract-feature-vector-for-each-detection-bbox

If it's not possible to share the checkpoints, can you shed some idea on how to obtain the proposal features?

-Vijay

INRIA

[Vijay Kumar]

Hi AVA team,

Can I assume that tensor operation "SecondStageBoxPredictor/AvgPool" in your .pb file which has the dimension (no_proposals, 1,1,1024) are the features?

-Vijay

INRIA