Update README.md

README.md

@@ -94,6 +94,7 @@ python2 evaluate_ate_scale.py ground_truth/kitti_ground_truth/00.txt sample/seq0
 We tried our best to improve the existing SVO, but this code is by no means perfect. That being said, we would like to point out some of the noticeable problems in our code:
 1. The original SVO is not designed to handle too much keyframes (KFs). Therefore, the system begins to slow down after accumulating too much KFs. For example, the system is running smoothly on Oxford Robotcar dataset because of the high frame-per-second (16 FPS); in contrast, it doesn't handle large amount of KFs well on KITTI dataset (10 FPS). (It is always better to use a high FPS camera)
 2. As mentioned in the paper, we use constant velocity model to handle extreme brightness condition (i.e., complete blank image), and then we back-project more points when new features can be observed. In practice, that makes sense for outdoor driving conditions. However, if your application requires proper relocalization, you would need to implement that by yourself.
+3. The network is trained to process depth maps with the image size of 256x512x3. There are two image resizing steps (downsizing the image and upsizing the depth map) in the VO pipeline. Although we don't see any noticeable problem doing it, it is definitely faster to process the images with the size of 256x512x3.
 
 We hope that you can further extend the functionality of this work, and make the existing SVO even better (which is an awesome piece of work).