If I understood the paper correctly, in the joint estimation model, the joint class prediction predicts which joint index a particular point belongs to (with 0 being not a member of a joint). So for a K-links system with K-1 joints, the joint classifier should have an output of shape (K, ) for each point.
When I checked the code for the architecture (Line-129) by default the joint_est_model() takes the argument of maximum parts to be 3 (i.e. K=3) and it is not being set to the current category's number of parts.
This may make sense for the eyeglasses category, but for others like ovens (K=2) and drawers (K=4), it doesn't. On inspecting the provided meta-graphs for these two categories, the output shape is always (3,). I changed the code to account for this but of course, since the graph now is different, the checkpoints won't load.
Assign requires shapes of both tensors to match. lhs shape= [1,128,2] rhs shape= [1,128,3]
[[Node: save/Assign_287 = Assign[T=DT_FLOAT, _class=["loc:@SPFN/joint_net/fc4_3/weights"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/device:GPU:0"](SPFN/joint_net/fc4_3/weights, save/RestoreV2/_575)]]
[[Node: save/RestoreV2/_582 = _Send[T=DT_FLOAT, client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_588_save/RestoreV2", _device="/job:localhost/replica:0/task:0/device:CPU:0"](save/RestoreV2:291)]]
****
Are the provided checkpoints and code correct only for the eyeglasses category while for others it isn't?
Edit: It still works for the oven (K=2) because the last logit is always 0, so when the arg-max is performed, one always get classification as {0, 1}. However, it seems more of an issue for drawers where K=4 where output should be in {0, 1, 2, 3} and not {0, 1, 2}.
If I understood the paper correctly, in the joint estimation model, the joint class prediction predicts which joint index a particular point belongs to (with 0 being not a member of a joint). So for a K-links system with K-1 joints, the joint classifier should have an output of shape (K, ) for each point.
When I checked the code for the architecture (Line-129) by default the
joint_est_model()takes the argument of maximum parts to be 3 (i.e. K=3) and it is not being set to the current category's number of parts.This may make sense for the eyeglasses category, but for others like ovens (K=2) and drawers (K=4), it doesn't. On inspecting the provided meta-graphs for these two categories, the output shape is always (3,). I changed the code to account for this but of course, since the graph now is different, the checkpoints won't load.
Are the provided checkpoints and code correct only for the eyeglasses category while for others it isn't?
Edit: It still works for the oven (K=2) because the last logit is always 0, so when the arg-max is performed, one always get classification as {0, 1}. However, it seems more of an issue for drawers where K=4 where output should be in {0, 1, 2, 3} and not {0, 1, 2}.