search

TiagoCortinhal / SalsaNext

Uncertainty-aware Semantic Segmentation of LiDAR Point Clouds for Autonomous Driving
MIT License
417 stars 102 forks source link

SalsaNet to SalsaNext, difference from addition op to concat. #38

Closed zeng-hello-world closed 3 years ago

zeng-hello-world commented 3 years ago

Hi, I have read both SalsaNet and SalsaNext, and they are great work. I noticed that you have change the skip connection in docoder from add to concat. Have you compared the performance between them, and what is the qualitative results.

Thanks.

TiagoCortinhal commented 3 years ago

Hello2 @zeyu-hello!

We compared but we do not have any ablative study done on that. The main idea was to try to improve the ability of the network to merge information coming from the decoder side. This will help us maintain more representational information coming from the encoder, that will later be merged through block IV.

I will try to find my runs with and without (I cannot assure I will find them) and if I find them I will share the differences! =)

zeng-hello-world commented 3 years ago

I know this concat is often applied in U-Net. The main idea I curious about this is that during deploymentation, concat op is not friendly for memory usage. If the error is somewhat acceptable, I could replace them to add and the inference speed will increase, too!

Thanks in advance! @TiagoCortinhal