Open bowang9711 opened 4 hours ago
Sparse4Dv2 initializes instance features as 0 and removes first two layers in single_frame_decoder, which improves detection performance. However, in our previous experiment, it's not helpful to online mapping, so we keep all layers for mapping.
Could you please clarify the rationale behind the inclusion of a [2:] slicing operation in the operation_order of det_head within the configuration, while it is absent in the map component?