Training own LiDAR model gives significant increase in IOU compared to the paper

[seamie6]

After training my own LiDAR model, I get an IOU score of 63.8. This is quite a bit higher than the IOU presented in the paper, 60.8. Perhaps it is a typo in the paper?? I used the same parameters as the camera+radar model:

   --exp_name="rgb_mine" \
   --dset='trainval' \
   --data_dir=XYZ \
   --device_ids=[0,1,2,3] \
   --ncams=6 \
   --batch_size=8 \
   --grad_acc=5 \
   --res_scale=2 \
   --do_rgbcompress=True \
   --max_iters=25000 \
   --log_freq=100 \
   --val_freq=100 \
   --save_freq=5000 \
   --nworkers=12 \
   --nsweeps=5 \
   --use_radar=False \
   --use_metaradar=False \
   --use_radar_filters=False \
   --use_lidar=True \

I am currently training the camera+radar model to see what happens there, again using the same parameters.

[seamie6]

C+R gives a result similar to the paper, 55.5