Under a single hard ego-side communication budget, gradient-boosted trees plus a multilayer perceptron predict each candidate block's gain and a greedy knapsack allocates bandwidth, reaching 99% of full-fusion AP@0.5 on OPV2V at 20.0 kB per frame

It introduces a selection mechanism that predicts a candidate feature block's gain before transmission, using an ensemble of gradient-boosted trees and a multilayer perceptron to estimate from pre-transmission metadata how many objects a candidate block would add to what the ego alone detects. Existing cooperative-perception work typically reports how many bytes were transmitted, whereas this turns 'what to share' into a pre-transmission gain-estimation problem, giving bandwidth allocation a benefit signal to act on. The method description names the model families (gradient-boosted trees and a multilayer perceptron ensemble) and the prediction target (added objects), and evaluates on the OPV2V benchmark with AP@0.5; the loaded text is incomplete and does not include ablation or predictor-accuracy detail tables.

It formulates budget allocation as a greedy knapsack across helpers, spatial blocks and numerical fidelity (fp16/int8/int4), jointly deciding what to share and at what precision under a single hard ego-side communication budget. Compared with a per-agent top-k baseline, the allocator brings the fidelity dimension into the decision space, yielding higher accuracy at equal budgets. On OPV2V the allocator reaches 99% of full-fusion AP@0.5 at 20.0 kB per frame (full fusion is 282.9 kB); with int4 it matches full-fusion accuracy at 10.0 kB where the per-agent top-k baseline needs 50.0 kB; at equal budgets of 10-50 kB it improves on that baseline by +0.3 to +0.9 AP points.

It models energy and latency without hardware, combining measured 5G modem power, a per-operation model calibrated on measured Jetson TX2 data, and replayed 5G driving traces, separating the radio energy that selection controls from the helpers' compute energy. Existing work reports transmitted bytes without their energy cost, whereas this brings energy into the evaluation and identifies the helpers' compute as dominating total energy. Across all 243 parameterisations tested, the allocator reaches 97% of full-fusion accuracy at lower radio energy than the baseline, and in 240 of 243 on total energy, which the text attributes to the helpers' compute dominating.

Sharing the helpers' box lists is a very low-cost complement, adding two to three AP points for under 1 kB. This suggests that beyond feature sharing, lightweight box-level information still yields sizeable accuracy gains at a bandwidth cost far below feature transmission. The text gives the quantitative description of 'under 1 kB' and 'two to three points', but the incomplete read does not include the full experimental setup for this combination.

The result targets cooperative-perception systems operating under a single hard ego-side communication budget, in settings where candidate-block gain can be predicted from pre-transmission metadata and the budget can be allocated discretely across helpers, spatial blocks and numerical fidelity; its energy conclusions rest on measured 5G modem power, a per-operation model calibrated on measured Jetson TX2 data, and replayed 5G driving traces, so it applies mainly to the vehicular and edge-IoT settings those models describe. For designers aiming to cut V2X bandwidth and radio energy in cooperative perception, it offers a reusable allocation approach and an accuracy-bandwidth-energy trade-off reference.

The loaded text is incomplete and lacks figures, ablations and statistical detail, so the predictor's own accuracy, the contribution of different metadata combinations, and the exact composition of the 243 parameterisations remain open questions; the energy conclusions depend on measured 5G modem power, a per-operation model calibrated on Jetson TX2, and replayed 5G driving traces, so whether they hold under other hardware or traces needs further verification; and the observation that the helpers' compute dominates total energy also suggests that total-energy gains may lie more on the compute side than the communication side.