search

chufengt / ALM-pedestrian-attribute

Code for the paper "Improving Pedestrian Attribute Recognition With Weakly-Supervised Multi-Scale Attribute-Specific Localization", ICCV 2019, http://arxiv.org/abs/1910.04562.
Apache License 2.0
188 stars 66 forks source link

Custom backbone instead of Inception #42

Closed sudapure closed 3 years ago

sudapure commented 3 years ago

From the code-base and paper implementation it looks like inception backbone is very native and adaptive to the complete attribute specified localization network, however the BNInceptionNet is bit heavy and contribute to high number of latency time during inference which makes it practically not suitable for the real-time surveillance scenarios, so will it be possible to replace the inception backbone network with any other lightweight backbones like mobilenet ,effecientNet-l0, etc, if it is feasible then how much it could impact the speed/accuracy factor.

chufengt commented 3 years ago

Sure. It's ok to replace the backbone network. The only modification is to choose a proper location to append the ALM modules. Speed can be improved with lightweight backbones.

nanhui69 commented 3 years ago

Sure. It's ok to replace the backbone network. The only modification is to choose a proper location to append the ALM modules. Speed can be improved with lightweight backbones.

in such case ? where can we use it for replace the model backbone..?

chufengt commented 3 years ago

Sure. It's ok to replace the backbone network. The only modification is to choose a proper location to append the ALM modules. Speed can be improved with lightweight backbones.

in such case ? where can we use it for replace the model backbone..?

BNInception class in inception_iccv.py defined the backbone. You can replace it with others, like ResNet, and modify the corresponding codes in the InceptionNet class.