[Feature]: Remove non-inference data from cooperative perception network payloads #191

Problem or motivation

The current cooperative perception payload contains data that is not required for model inference.

For IntermediateFusionDataset, every CAV serializes several fields into Entity.auxillary, including:

• object_ids
• object_bbx_center
• infra
• velocity
• time_delay
• spatial_correction_matrix
• voxel_num_points
• voxel_features
• voxel_coords

The payload is serialized with pickle, stored as Protobuf bytes, and then converted to JSON/Base64 by CosimService. This introduces additional serialization overhead.

During a 100-CAV Town03 scenario, one CAV generated a payload of 382,226 bytes, while the Vanetza GeoNetworking maximum SDU size is 1,398 bytes:

sending cosim payload from vehicle cav-66 on channel 180, payload bytes=382226
Model error: GN-Data.request rejected

Some transmitted fields, particularly ground-truth boxes and object IDs, are used for dataset labels, evaluation, or visualization but are not consumed by point_pillar_where2comm.forward() during inference.

Transmitting these fields through the simulated ITS-G5 channel wastes bandwidth and makes the network traffic less representative.

Proposed solution

Define an explicit, model-specific wire payload for cooperative perception and remove fields that are not required for remote inference.

For the current Where2comm intermediate-fusion pipeline:

1. Audit every field currently written by IntermediateFusionDataset.extract_data().
2. Classify fields as:
   • required remote inference input;
   • synchronization or transformation metadata;
   • ground truth;
   • evaluation or visualization-only data.
3. Keep ground truth, labels, object IDs, and visualization-only data local or transfer them outside the simulated radio channel.
4. Transmit only the fields required to reconstruct the remote inference input.
5. Add per-field and total serialized payload-size logging.
6. Add tests verifying that removing non-inference fields does not change model predictions.

Until distributed feature extraction is implemented, the following voxel tensors may still be required:

• voxel_features
• voxel_coords
• voxel_num_points

Therefore, this issue is expected to reduce the payload size but may not be sufficient by itself to fit the payload into a single GeoNetworking SDU.

Area

Cooperative perception

Alternatives considered

No response

Additional context

No response