Open MariasStory opened 5 years ago
Thanks for posting this! The model indeed is very big in size and complex:
Figure 2: Our DSFD framework uses a Feature Enhance Module (b) on top of a feedforward VGG/ResNet architecture...
WIDERFace_DSFD_RES152.pth (459M)
But I will take a look at the paper, sometimes you can apply some of the ideas when training an own model. Also always good to have a reference of state of the art methods and curious how close you can get to state of the art with smaller and faster models, that are actually web friendly.
To answer the following question:
What do we need to get following performance? https://github.com/TencentYoutuResearch/FaceDetection-DSFD
I can't really tell you to be honest. First we would have to figure out, where the current models of face-api.js (ssd mobilenet and tiny face detector) fit into the picture performancewise.
Another reference for a good benchmark: https://github.com/deepinsight/insightface/tree/master/RetinaFace
The tensorflow implementation: https://github.com/610265158/DSFD-tensorflow
Hi, your solution is cool!
What do we need to get following performance? https://github.com/TencentYoutuResearch/FaceDetection-DSFD
DSFD achieves the best performance among all of the state-of-the-art face detectors based on the average precision (AP) across the three subsets, i.e., 96.6% (Easy), 95.7% (Medium) and 90.4% (Hard) on validation set, and 96.0% (Easy), 95.3% (Medium) and 90.0% (Hard) on test set.
One problem that I see is the model size: WIDERFace_DSFD_RES152.pth (459M) https://drive.google.com/uc?id=1WeXlNYsM6dMP3xQQELI-4gxhwKUQxc3-&export=download
Would be nice to get closer to this one ;-). Even if it takes more memory and time, there might be a use case for it.
Thanks.