Anchor size in the config file

[Blade6570]

Hey,

How did you generate the anchor sizes and anchor bottom heights values for this dataset. How do they affect the accuracy the final accuracy? If you have some documentation or information please share it.

[mpitropov]

Hi, I'm not sure of the change to the final accuracy as I didn't train a model with the KITTI values to compare my results. I decided to calculate them since our Lidar is placed high up on the vehicle.

This is the code I used:

# Calculate anchor head bottom statistics
car_list = []
ped_list = []
pickup_truck_list = []
for gt_dict in gt_dicts:
    for i in range(len(gt_dict['annos']['gt_boxes_lidar'])):
        if gt_dict['annos']['name'][i] == 'Car':
            car_list.append(gt_dict['annos']['gt_boxes_lidar'][i])
        if gt_dict['annos']['name'][i] == 'Pedestrian':
            ped_list.append(gt_dict['annos']['gt_boxes_lidar'][i])
        if gt_dict['annos']['name'][i] == 'Pickup_Truck':
            pickup_truck_list.append(gt_dict['annos']['gt_boxes_lidar'][i])
car_mean = np.mean(car_list, axis=0)
print('mean car', car_mean)
print('bottom car', car_mean[2] - car_mean[5]/2)
ped_mean= np.mean(ped_list, axis=0)
print('mean ped', ped_mean)
print('bottom ped', ped_mean[2] - ped_mean[5]/2)
pickup_truck_mean = np.mean(pickup_truck_list, axis=0)
print('mean pickup truck', pickup_truck_mean)
print('bottom pickup truck', pickup_truck_mean[2] - pickup_truck_mean[5]/2)