#!python train.py --task=cls
!python eval_cls.py
0.9758656873032528
The models seem to detect anything that has a significant sitting area is a chair. The vase which is wrongly classified as a lamp, is confusing for humans as well.
The lamp which is wrongly classified as a vase is probably due to us maxpooling and not neighbour pooling which means use of nearby points information.
#!python train.py --task=seg
The good examples are really good as you can see.
The bad examples are not really bad since they are confusing.
In the first example the back of chair is identified as the pillow which is a reasonable mistake that could be made by humans as well. Plus armrest and legs are the same thing in gt which means its confusing for humans as well.
The second example, there are no legs in gt but there are legs in predicted sofa. This is probably due to model learning some part of the model is always legs.
The third example, the armrest could be used for sitting and could be misunderstood by the network to be seating area.
Top performing models are sofas and low performing models are sofas, couches etc.
# !python eval_seg.py --i 351 --exp_name 'b1'
# !python eval_seg.py --i 26 --exp_name 'b2'
# !python eval_seg.py --i 61 --exp_name 'b3'
# !python eval_seg.py --i 297 --exp_name 'g1'
# !python eval_seg.py --i 471 --exp_name 'g2'
# !python eval_seg.py --i 562 --exp_name 'g3'
!$CUB_HOME