Van and Cyclist

[freshwk]

Hi,if i want to train and test Van and Cyclist category ,what do I need to modify in the BAT_Car.yaml? I am so sorry to bother you again! I am following your work and I need to do a complete experiment to test my idea.It would be better if you have time to upload the config file to github.Thank you!

[Ghostish]

You need to change the category_name in the BAT_Pedestrian.yaml to Cyclist / Van.

[freshwk]

You need to change the category_name in the BAT_Pedestrian.yaml to Cyclist / Van.

Unfortunately，i get a much bad result in Van category 32.85/39.88 after retraining your code .I don't know what went wrong.My configs are as follows.

data

dataset: kitti path: /2TB/kitti/training category_name: Van # [Car, Van, Pedestrian, Cyclist, All] search_bb_scale: 1.25 search_bb_offset: 2 model_bb_scale: 1.25 model_bb_offset: 0 template_size: 512 search_size: 1024 random_sample: False sample_per_epoch: -1 degrees: True # use degrees or radians box_aware: True num_candidates: 4 coordinate_mode: velodyne up_axis: [0,0,1]

model configuration

net_model: BAT use_fps: True normalize_xyz: False feature_channel: 256 #the output channel of backbone hidden_channel: 256 #the hidden channel of xcorr out_channel: 256 #the output channel of xcorr vote_channel: 256 #the channel for vote aggregation num_proposal: 64 k: 4 use_search_bc: False use_search_feature: False bc_channel: 9

loss configuration

objectiveness_weight: 1.5 box_weight: 0.2 vote_weight: 1.0 seg_weight: 0.2 bc_weight: 1.0

testing config

reference_BB: previous_result shape_aggregation: firstandprevious use_z: False limit_box: True IoU_space: 3

training

batch_size: 50 #batch_size per gpu workers: 10 epoch: 60 from_epoch: 0 lr: 0.001 optimizer: Adam lr_decay_step: 12 lr_decay_rate: 0.2 wd: 0

[Ghostish]

Hello, the configuration you used to train your model is just OK. But you should use a smaller batch size (e.g. batch_size = 24 for a single GPU machine).

I just re-trained a new Van model using your configuration with the batch_size = 24. Here is the results I got:

By the way, since the training samples of the Van and Cyclist in the KITTI dataset are not sufficient for stable training, the test results may have a great variance across different runs. To test your model, you should focus on the Car/Pedestrian instead.