XG comments

Results 14 comments of

XG

Inference on single stereo image pairs

> Why is my prediction so bad？ ![image](https://user-images.githubusercontent.com/44777088/97152814-70d8c400-17ac-11eb-829b-93da9c7ac5d6.png) Compare to the paper result, looks like this is actually the best result AnyNet can get.

Some error about spn

I think is your environment problem, I compile with no such errors. However I couldn't found a way to install spn_t1 properly, the make.sh under spn_t1 folder doesn't install it...

Add the kitti evaluation code

@maudzung Hello, I found that your ImageSets files are not following the conventions in KITTI 3D, which often use 3712 pointcloud bin as train-split and 3769 as val-split. The split...

What is your IOU threshold for 'EASY /MODERATE /HARD' AP?

I doubt on it too. Since I can't reproduce the pedestrian paper result on KITTI train-val split. The log shows below. step=296950, steptime=0.267, cls_loss=0.0828, cls_loss_rt=0.0787, loc_loss=0.438, loc_loss_rt=0.317, rpn_acc=0.999, prec@10=0.319, rec@10=0.983,...

Must be sparse before gdc？

I see the --disable_subsample param under main_batch.py, have you ever try it?

Where are the inputs and outputs of AnyNet

Looks like the author does not release the code related to disparity map/depth map generation.

超过100MB文件下载时报禁止访问，但官方客户端能够正常下载

> > 尝试了可以把大文件切片，可以从linux的命令行端下载，再把切片拼接起来。 > > > > * 切片：`split -b 100M large_file.txt large_file_part_` > > * 拼接：`cat large_file_part_* > reassembled_large_file.txt` > > 速度有点慢呀在网页端下载也是这个速度，是阿里云盘的问题。你在命令行把下载并发量和缓存开满，并发总量能到7MB/s。 aliyunpan config set -max_download_parallel 20 aliyunpan...

XG