Low performance in multiclass detection

[smell-fishy]

Hi, I wanted to train a pointpillar model to classify 3 class and I modified the configure as:

model: { second: { voxel_generator { point_cloud_range : [0, -39.68, -3, 69.12, 39.68, 1] voxel_size : [0.16, 0.16, 4] max_number_of_points_per_voxel : 100 } num_class: 3 ... target_assigner: { anchor_generators: { anchor_generator_stride: { sizes: [1.6, 3.9, 1.56] # wlh strides: [0.32, 0.32, 0.0] # if generate only 1 z_center, z_stride will be ignored offsets: [0.16, -39.52, -1.78] # origin_offset + strides / 2 rotations: [0, 1.57] # 0, pi/2 matched_threshold : 0.6 unmatched_threshold : 0.45 } } anchor_generators: { anchor_generator_stride: { sizes: [0.6, 1.76, 1.73] # wlh strides: [0.16, 0.16, 0.0] # if generate only 1 z_center, z_stride will be ignored offsets: [0.08, -19.76, -1.465] # origin_offset + strides / 2 rotations: [0, 1.57] # 0, pi/2 matched_threshold : 0.5 unmatched_threshold : 0.35 } } anchor_generators: { anchor_generator_stride: { sizes: [0.6, 0.8, 1.73] # wlh strides: [0.16, 0.16, 0.0] # if generate only 1 z_center, z_stride will be ignored offsets: [0.08, -19.76, -1.465] # origin_offset + strides / 2 rotations: [0, 1.57] # 0, pi/2 matched_threshold : 0.5 unmatched_threshold : 0.35 } } ... } } }

train_input_reader: { ... class_names: ["Car", "Cyclist", "Pedestrian"] ... database_sampler { ... sample_groups { name_to_max_num { key: "Car" value: 15 } } sample_groups { name_to_max_num { key: "Pedestrian" value: 6 } } sample_groups { name_to_max_num { key: "Cyclist" value: 6 } } database_prep_steps { filter_by_min_num_points { min_num_point_pairs { key: "Car" value: 5 } min_num_point_pairs { key: "Pedestrian" value: 10 } min_num_point_pairs { key: "Cyclist" value: 10 } } } ... }

...

eval_input_reader: { ... class_names: ["Car", "Pedestrian", "Cyclist"] ... }

BUT the validation performance in cyclist and pedestrian detection is poor as follow:

Car [email protected], 0.70, 0.70: bbox AP:90.77, 87.93, 83.59 bev AP:89.98, 82.29, 81.22 3d AP:81.93, 69.87, 67.45 aos AP:90.58, 87.33, 82.54 Car [email protected], 0.50, 0.50: bbox AP:90.77, 87.93, 83.59 bev AP:90.85, 89.98, 89.45 3d AP:90.85, 89.80, 89.13 aos AP:90.58, 87.33, 82.54 Cyclist [email protected], 0.50, 0.50: bbox AP:0.36, 16.96, 15.94 bev AP:0.29, 13.38, 12.91 3d AP:0.28, 12.32, 12.30 aos AP:0.00, 15.15, 14.42 Cyclist [email protected], 0.25, 0.25: bbox AP:0.36, 16.96, 15.94 bev AP:0.31, 15.23, 14.56 3d AP:0.31, 15.23, 14.56 aos AP:0.00, 15.15, 14.42 Pedestrian [email protected], 0.50, 0.50: bbox AP:0.17, 4.70, 4.68 bev AP:0.02, 1.14, 1.14 3d AP:0.01, 1.14, 1.14 aos AP:0.08, 3.48, 3.47 Pedestrian [email protected], 0.25, 0.25: bbox AP:0.17, 4.70, 4.68 bev AP:0.07, 1.21, 1.21 3d AP:0.07, 1.21, 1.21 aos AP:0.08, 3.48, 3.47

Car coco [email protected]:0.05:0.95: bbox AP:70.46, 65.58, 64.50 bev AP:68.30, 64.00, 63.20 3d AP:56.65, 51.51, 50.05 aos AP:70.32, 65.17, 63.69 Cyclist coco [email protected]:0.05:0.70: bbox AP:0.36, 16.02, 15.07 bev AP:0.27, 13.03, 12.55 3d AP:0.27, 12.37, 12.01 aos AP:0.03, 14.44, 13.74 Pedestrian coco [email protected]:0.05:0.70: bbox AP:0.23, 3.95, 3.95 bev AP:0.02, 0.75, 0.75 3d AP:0.02, 0.71, 0.72 aos AP:0.11, 2.48, 2.48

[kaancolak]

Also i trained and got same result. Does anyone share correct configuration file?

[mickeyouyou]

so good result to hard to get it