Training on my own dataset about pointnet

Hi @charlesq34,
Thank you for your work. The code is really helpful.
What is the faceid field I see in the .h5 file? I have attached an image for your reference. I'm using a program called vitables to view the .h5 file.

I understand that label value at location (0,0) is the label for data at the same location and the normal value corresponds to the value returned from LIDAR(as I understand). I do see these fields being added to the .h5 file in the data_prep_util.py file but I don't see the faceid field there

Also if I pass my data to the helper function provided in the data_prep_util.py file, the generated .h5 file can be used for training right?

[charlesq34]

Hi @skymaze

You can set a MAX number of points during training, if there are less than MAX points, just replicate existing points should be fine -- since we use max that will not alter final result. If there are more than MAX, you can randomly subsample the points.

during test you can have a smaller batch size and even larger MAX to evaluate all points :)
or if there are too many you can compute result in two steps -- by firstly extracting point embedding and then aggregate.

Hope it helps!

Best,
Charles

[EndingCredits]

When batching presumably you need to mask the batch before pooling to ensure the extra missing points are not included?

I did something like this:

# Turns the input list into a single tensor, padding the outputs.
# Produces a mask of size batch_size x max_list_len x 1
def batchToArrays(input_list):
    max_len = 0
    out = []; seq_len = []; masks = []
    for i in input_list: max_len = max(len(i),max_len)
    for l in input_list:
        # Zero pad output
        out.append(np.pad(np.array(l,dtype=np.float32), ((0,max_len-len(l)),(0,0)), mode='constant'))
        # Create mask... (max_len x 1 vector with len(l) ones and rest zeros)
        # Extra dimension is left in for easy broadcasting
        masks.append(np.pad(np.array(np.ones((len(l),1)),dtype=np.float32), ((0,max_len-len(l)),(0,0)), mode='constant') )
return out, masks

and then:

def mask_and_pool(elems, mask, pool_type='max_pool'):
    # Use broadcasting to multiply
    masked_elems = tf.multiply(elems, mask)
    # Pool using max pooling
    embed = tf.reduce_max(masked_elems, axis=1)
return embed

My batch and num_points are set to None in the placeholder so I don't have to worry about the exact size, just about making sure everything has the same dimensions.

[charlesq34]

Hi @EndingCredits

If that's max pooling then it won't matter in fact. Replicating points will not alter max pooled result.
If it's average pooling then you do need to add a mask.