PaperlayerView exact revision on arXiv

Source-linked AI summary

PEARL: A Lightweight Prompt-based Feature Interpreter Framework for Real-Time, Anonymous, and Heterogeneous Collaborative Perception

Armin Maleki, Hayder Radha

arXiv:2609.27123v1cs.CVcs.LG

TL;DR

Heterogeneous sensors, models, and training data create feature-domain gaps for collaborative perception, while existing alignment approaches can require configuration disclosure and struggle with real-time adaptation to unseen agents. PEARL combines sparse detection and dense domain-invariant prompt interpreters with low-rank prompts to select interpreters for new agents. It reports improvements over random selection and state-of-the-art heterogeneous CP frameworks, alongside millisecond-scale scoring latency.

  • Problem

    Heterogeneous CP must align features across agents despite differing sensors, models, and training data, while existing approaches can require configuration access and struggle with unseen agents and real-time deployment.

    Source: §I · p. 2, F. Real-Time Implementation · p. 6

  • Method

    PEARL uses parallel lightweight LWSD and LWDDI prompt interpreters, with a two-stage adaptation strategy that updates lightweight modules while freezing complex perception components.

    Source: §I · p. 2

  • Results

    3.3% AP@0.5 and 5.6% AP@0.7 gains over state-of-the-art heterogeneous CP frameworks are reported for offline training on average.

    Source: §I · p. 3

  • Takeaways & Limitations

    PEARL supports real-time assignment of unseen agents without sensor or model metadata disclosure, with 1.67 ms average similarity-scoring latency.

    Source: §I · p. 3, §V · p. 11

Abstract

from arXiv · show

Heterogeneity across Collaborative Perception (CP) agents is a major challenge for emerging CP frameworks due to domain gaps from differing sensors, architectures, and training data. Prior works mitigate this challenge by aligning features in a unified space via model retraining or per-agent-type interpreters. These strategies (a) require access to neighbor configurations, (b) do not fully address real-time CP deployment, and (c) generalize poorly to unseen agents joining at run time. To overcome these challenges, we present PEARL, a Prompt-Embedding framework for Anonymous and Real-time Lightweight heterogeneous CP. PEARL supports multiple CP interpreters and selects one for a new-joining agent in real time using two lightweight, multi-scale interpreters trained in parallel: a sparse-detection (LWSD) interpreter that aligns salient regions for cooperative detection, and a dense, domain-invariant (LWDDI) interpreter that produces agent-invariant features for fast interpreter selection. Both interpreters use low-rank visual prompts to reduce computation, storage, and model complexity. Extensive experiments on simulated (OPV2V, V2XSet) and real (DAIR-V2X) datasets show that PEARL generalizes across simulated and real-world cooperative driving scenarios. Its real-time model-selection strategy yields an 8.2% Average Precision (AP) gain over a random-selection baseline while running in 1.67 ms on average. Although primarily designed for real-time CP, PEARL also outperforms state-of-the-art heterogeneous CP frameworks under traditional offline training by 5.6% AP on average while reducing communication cost by up to 34.7 times. Equally important, PEARL does not require sharing agents' configurations or model settings, thereby protecting information that may be proprietary or private. These results establish PEARL as a scalable and practical framework for heterogeneous collaborative perception.

I. INTRODUCTION

PEARL addresses heterogeneous collaborative perception with lightweight, anonymous interpreter selection for new agents. It combines sparse prompts for detection alignment with dense, domain-invariant features for selection, and reports gains over random selection and state-of-the-art heterogeneous CP frameworks.

Motivation and prior work

  • Prior approaches retrain models or use per-agent-type interpreters, but struggle to balance generalization to unseen agents, accuracy, privacy, and real-time feasibility.Retraining can be slow and risk privacy or single-vehicle performance; per-type interpreters add complexity and may reduce performance.Source: §I · p. 1

Runtime selection

  • Runtime selection based on domain-invariant intermediate features avoids running a full detector and does not require disclosure of a joining agent’s sensor or model metadata.The approach leverages shared semantics across model and weight variants to assign an available model to an unseen agent.Source: §I · p. 2

Framework and interpreters

  • PEARL uses parallel LWSD and LWDDI pipelines: sparse prompts align detection-critical features, while dense prompts learn agent-invariant representations for anonymous, real-time model assignment.Low-rank multi-scale prompts reduce computation, and selection uses an aggregated similarity score on LWDDI features.Source: §I · p. 2

Adaptation to unseen agents

  • A two-stage strategy adapts to unseen heterogeneous agents by updating lightweight modules while freezing complex perception components.This reduces retraining cost and is intended to preserve single-agent performance.Source: §I · p. 2

Experiments and contributions

  • 1.67 ms average similarity-scoring latency supports PEARL’s real-time interpreter assignment strategy.The strategy is designed to scale across multiple candidate interpreters and joining agents.Source: §I · p. 3
  • 3.3% AP@0.5 and 5.6% AP@0.7 gains over state-of-the-art heterogeneous CP frameworks are reported for offline training on average.For real-time selection, PEARL gains 5.3% AP@0.5 and 8.2% AP@0.7 over random selection on average.Source: §I · p. 3

II. RELATED WORK

The paper situates PEARL among collaborative-perception fusion strategies, heterogeneous feature-alignment methods, and visual-prompt approaches. Its methodology uses multi-scale interpreters and a detection pipeline to align and fuse neighbor features with ego features.

A. Collaborative Perception

  • Collaborative perception commonly shares raw data, detections, or intermediate features, trading accuracy and communication bandwidth across early, late, and intermediate fusion.Intermediate fusion shares latent features and balances the accuracy of early fusion with the bandwidth efficiency of late fusion.Source: A. Collaborative Perception · p. 3

B. Heterogeneous CP

  • Heterogeneous CP methods retrain models into a shared semantic space or use interpreters, but may require known configurations, metadata, or joint retraining.PEARL is positioned as an anonymous, real-time interpreter for newly joining agents.Source: B. Heterogeneous CP · p. 3

C. Visual Prompts

  • Visual prompts are small learnable tensors for adapting intermediate features; prior work reports sparse prompts as better suited to detection gaps than dense prompts.Prompting has also been applied to heterogeneous CP in PolyInter and Faster-HEAL [24].Source: C. Visual Prompts · p. 3

III. METHODOLOGY

  • PEARL has a detection pipeline for ego-space alignment and an adversarial domain-classification pipeline for domain-invariant BEV features.Stage-1 trains neighbor-side modules while keeping encoders and the ego perception stack frozen; Stage-2 fine-tunes lightweight prompts, the classifier, and the new agent’s compressor and multi-scale extractor.Source: §III · p. 3, Figure 1 · p. 4

A. Pipeline-1: Detection

  • The detection pipeline attaches a visual prompt to each neighbor type and scale, then uses channel-selection and spatial-transformer modules to refine multi-scale BEV features.Two multi-head cross-attention branches align neighbor channels to ego channels, including a prompt-guided value stream.Source: A. Pipeline-1: Detection · p. 4
  • LWSD prompts emphasize detection-relevant regions, and the same foreground estimator is reused to support lightweight fusion.The estimator gates a learnable prompt and predicts spatial objectness probabilities used as normalized per-agent fusion weights.Source: A. Pipeline-1: Detection · pp. 4–5
  • After channel alignment, 3D spatial attention aligns neighbor features to ego features at each scale.The attention operates on features that LWSD has already made salient in detection-relevant regions.Source: A. Pipeline-1: Detection · p. 4
  • Foreground-weighted features are fused per scale, then upsampled by the ego multi-scale decoder, concatenated, and sent to the detection head.Source: A. Pipeline-1: Detection · pp. 4–5

B. Pipeline-2: Adversarial Domain Classification

The adversarial domain-classification pipeline uses dense prompts to make heterogeneous neighbor BEV features comparable with ego features. A shared classifier and adversarial training produce domain-invariant features for real-time model selection.

B. Pipeline-2: Adversarial Domain Classification

  • Dense prompts are attached at each scale and neighbor type to align neighbor BEV features to ego space before classification.The pipeline takes multi-scale ego and compressed-neighbor BEV features.Source: B. Pipeline-2: Adversarial Domain… · p. 5
  • The resulting domain-invariant features are intended to be comparable across heterogeneous vehicles and support fast real-time model selection.The pipeline uses lightweight prompts and a shallow classifier.Source: B. Pipeline-2: Adversarial Domain… · p. 5
  • Adversarial training makes domains indistinguishable through prompts while the shared classifier minimizes domain-classification error.A gradient-reversal layer realizes the min–max optimization.Source: B. Pipeline-2: Adversarial Domain… · p. 5

C. Low-Rank Visual Prompts

PEARL factorizes its visual prompts into low-rank components to reduce the cost of learning full 3D prompt tensors. The reduced parameterization is intended to preserve accuracy while making adaptation and runtime use feasible.

C. Low-Rank Visual Prompts

  • Full 3D prompts can become impractical for high-resolution BEVs or many agents and may slow convergence and encourage overfitting.Source: C. Low-Rank Visual Prompts · p. 5
  • PARAFAC factorization replaces each full prompt tensor with three 2D factors for channel, height, and width, using decomposition rank R.The reconstructed prompt value is indexed by channel c and spatial location (h, w).Source: C. Low-Rank Visual Prompts · p. 5
  • R(C + H + W) parameters per prompt reduce computation overhead while maintaining high accuracy in both pipelines, making Stage-2 and real-time implementation feasible.Source: C. Low-Rank Visual Prompts · p. 5

D. Stage-1: Heterogeneous Base Training

Stage-1 trains both PEARL pipelines around the ego BEV feature space, mapping heterogeneous agents into a shared reference while keeping heavy components frozen. Its detection and domain-invariance objectives combine cooperative supervision with alignment losses.

D. Stage-1: Heterogeneous Base Training

  • Stage-1 maps heterogeneous agents to the ego BEV reference space while jointly training both pipelines, neighbor compressors, multi-scale extractors, and foreground estimators with encoders and the ego detection stack frozen.Neighbors transmit intermediate BEV features without model or sensor metadata.Source: D. Stage-1: Heterogeneous Base Training · p. 5
  • Pipeline-1 combines collaborative detection with per-neighbor detection and semantic-alignment objectives, including matching feature means and variances to the ego.Per-neighbor box classification and regression losses provide auxiliary supervision.Source: D. Stage-1: Heterogeneous Base Training · p. 5
  • Pipeline-2 learns domain-invariant features using adversarial training and moment matching between prompted neighbor and ego features.Its total loss combines the Pipeline-2 objectives during end-to-end Stage-1 training.Source: D. Stage-1: Heterogeneous Base Training · pp. 5–6

E. Stage-2: New Heterogeneous Agent Training

Stage-2 adapts a limited set of modules to integrate a new heterogeneous agent into the shared base. At runtime, PEARL selects a pretrained interpreter from BEV features using domain-invariant similarity, without requiring configuration metadata.

E. Stage-2: New Heterogeneous Agent Training

  • Stage-2 unfreezes only LWSD, the foreground estimator, LWDDI, the domain classifier, and the new agent’s multi-scale extractor and compressor; other components remain frozen.Interpreters build on a shared base so the ego can host many while storing one base module.Source: E. Stage-2: New Heterogeneous Agent… · p. 6

F. Real-Time Implementation

  • The ego may encounter many agent types, including unseen types, while nominally familiar agents can produce different BEV semantics after changes in data or training.Selecting an interpreter from metadata can compromise privacy; exhaustive real-time testing or retraining is infeasible.Source: F. Real-Time Implementation · p. 6
  • For a joining agent, the ego evaluates compressed BEVs through pretrained LWDDI paths and selects the interpreter whose domain-invariant features have the highest cosine similarity to its own.This assignment is performed at runtime using only BEV features.Source: F. Real-Time Implementation · p. 6
  • Selection takes 1.67 ms and shares only BEV features; one stored model can serve multiple agents whose similarity exceeds a threshold.Stage-2 adaptation can spawn interpreters from one heterogeneous base with negligible storage.Source: F. Real-Time Implementation · p. 6

IV. EXPERIMENTAL RESULTS

PEARL is evaluated on simulated and real collaborative-perception datasets under heterogeneous base training and new-agent adaptation. It outperforms PolyInter across the reported settings, including real-world DAIR-V2X adaptation.

A. Dataset and Experimental Setting

  • Evaluation covers OPV2V, V2XSet, and DAIR-V2X, with encoder families and voxel variants differing across ego and neighbor agents and unseen agents introduced one at a time for adaptation.Multi-modal experiments also include ResNet and EfficientNet encoders.Source: A. Dataset and Experimental Setting · p. 6

B. Performance Comparison

  • Across OPV2V, V2XSet, and DAIR-V2X, PEARL outperforms PolyInter in heterogeneous base training and new-agent adaptation, including LiDAR- and camera-agent settings.The evaluation uses DAIR-V2X infrastructure-ego and vehicle-neighbor samples to represent real-world vehicle-to-infrastructure perception.Source: B. Performance Comparison · p. 7
  • On OPV2V, PEARL’s Stage-1 gains over PolyInter are +2.4/+4.5% and +2.3/+5.5% AP@0.5/@0.7 for pp4-vn4 and vn4-sd2, respectively.On DAIR-V2X, Stage-1 gains include +6.2/+8.2% and +5.3/+7.8% under the same settings.Source: B. Performance Comparison · p. 7
  • On OPV2V Stage-2 adaptation, PEARL’s gains over PolyInter reach +4.0/+6.2% AP@0.5/@0.7 for LiDAR vn6 and are +1.7/+2.2% for ResNet and +3.1/+3.3% for EfficientNet camera agents.For LiDAR adaptation, the largest reported improvements occur on vn6, which has longer range and coarser voxelization.Source: B. Performance Comparison · p. 7
  • On DAIR-V2X Stage-2 adaptation under pp4-vn4, gains are +5.8/+7.7% (pp4), +6.2/+7.7% (sd1), and +6.4/+7.6% (vn6) AP@0.5/@0.7.V2XSet also shows adaptation gains across pp4, sd1, and vn6 under both reported heterogeneous settings.Source: B. Performance Comparison · p. 7

C. Model efficiency and communication cost

PEARL reduces adaptation and communication costs, then selects interpreters for anonymous agents using domain-invariant features. Selection scoring averages 1.67 ms per sample, while full selection averages 48.67 ms; parallel evaluation and one-time assignment help keep deployment practical.

C. Model efficiency and communication cost

  • In Stage-1, PEARL uses 1.1M–4.4M fewer trainable parameters than PolyInter across heterogeneous bases.Stage-2 adapts a compact model of approximately 4.1M parameters; PEARL's total parameter count is lower in most scenarios.Source: C. Model efficiency and communication… · pp. 7–8
  • PEARL transmits compressed BEV features of size 64 × 64 × 128 (∼0.52M elements), reducing feature size versus PolyInter by 8.7×–34.7×, approximately 24× on average.Under fixed precision, communication cost scales linearly with transmitted elements.Source: C. Model efficiency and communication… · p. 8, Table IV · p. 8

D. Real-Time Implementation

  • 1.67 ms per sample is the average similarity-scoring time, while end-to-end interpreter selection averages 48.67 ms and runs only once when an agent joins.After assignment, the ego uses only the selected interpreter; selection is significantly cheaper than full collaborative inference.Source: D. Real-Time Implementation · p. 9
  • For k = 25 candidate interpreters, concurrent evaluation requires approximately 44 GB of memory and yields latency comparable to a single-interpreter evaluation.GPU memory grows approximately linearly with candidate interpreters and joining agents, while parallelism bounds latency by hardware and memory availability.Source: D. Real-Time Implementation · p. 9, Figure 4 · p. 10
  • For multiple joining agents, selection can run in parallel with total memory scaling as k × n; (k = 13, n = 2) is handled on one GPU.Larger settings can be distributed across multiple GPUs with near-linear throughput scaling.Source: D. Real-Time Implementation · p. 9
  • Across OPV2V, V2XSet, and DAIR-V2X, higher domain-invariant-space similarity is consistently associated with stronger collaborative detection performance.Figure 3 compares average performance across candidate interpreters with matched-type interpreters for known agents and the highest-similarity selection for unseen pp6.Source: D. Real-Time Implementation · p. 9, Figure 3 · p. 10

E. Cross-Dataset Tests

Cross-dataset tests show that dataset shift affects interpreter compatibility even for the same nominal encoder type. In selection-baseline comparisons, PEARL outperforms strategies that omit domain-invariant features or replace cosine similarity with moment matching.

E. Cross-Dataset Tests

  • 95.4/83.7% AP@0.5/0.7 with Savg = 0.8746 for OPV2V-trained interpreters compares with 87.7/60.4% AP@0.5/0.7 and Savg = 0.8420 for V2XSet-trained interpreters.The comparison uses the top-3 selected interpreters for a pp4 encoder trained on OPV2V.Source: E. Cross-Dataset Tests · p. 10, Table V · p. 10
  • Even for the same nominal pp4 encoder type, dataset-level shift affects compatibility, so sensor or model metadata alone is insufficient to determine interpreter compatibility.PEARL uses feature-based similarity selection to identify compatible interpreters across intra- and cross-dataset variations.Source: E. Cross-Dataset Tests · p. 10

F. Model Selection Baseline

  • PEARL improves over B1 and B2 by +3.8/+4.8% in AP@0.5 and +6.9/+8.4% in AP@0.7, respectively.B1 uses cosine similarity without LWDDI; B2 uses LWDDI with moment-matching distance instead of cosine similarity.Source: F. Model Selection Baseline · p. 10, Table VI · p. 10

G. Ablation Study

The ablations test PEARL's multi-scale extraction, alignment, prompting, and fusion choices, and the conclusion describes its two-interpreter framework for anonymous heterogeneous collaboration. Results highlight multi-scale extraction as the largest tested contributor to detection performance.

G. Ablation Study

  • Removing the multi-scale extractor reduces AP by 9.3/19.3% at AP@0.5/AP@0.7 in Stage-1 and by 10.8/19.2% on average in Stage-2.The authors identify coarse-to-fine processing as important for heterogeneous feature alignment, particularly for vn6 with a larger domain gap.Source: G. Ablation Study · p. 11, Table VII · p. 11
  • Removing cross-attention lowers AP by 5.8% and 2.2% at AP@0.5, and by 3.5% and 8.0% at AP@0.7 in Stage-1 and Stage-2, respectively.The authors associate this effect with channel-wise semantic alignment across heterogeneous encoders.Source: G. Ablation Study · p. 11
  • A shared frozen extractor reduces AP by 1.0/3.7% in Stage-1 and 1.6/4.7% in Stage-2 on average; removing spatial attention reduces it by 1.6/3.9% and 1.0/2.5%, respectively.Removing the foreground estimator also causes a modest drop.Source: G. Ablation Study · p. 11, Table VII · p. 11
  • Using a single LWSD prompt across scales reduces performance by 3.6/8.2% in Stage-1 and 2.6/6.8% in Stage-2 on average.The comparison tests scale-specific prompting against one shared prompt.Source: G. Ablation Study · p. 11, Table VII · p. 11
  • Replacing learned foreground-weighted fusion with fixed 0.5 weights reduces performance by 1.1/1.3% in Stage-1 and 1.4/1.5% in Stage-2 on average.The authors state that adaptive spatial fusion better exploits complementary information from neighboring agents.Source: G. Ablation Study · p. 11, Table VII · p. 11
  • PARAFAC ranks R = [8, 16, 32] offer the best reported trade-off at approximately 30K parameters, with prompt weights of approximately 12K–50K versus approximately 1.83M for dense 3D prompts.The authors report comparable performance across settings for PARAFAC prompts; larger rank can slow convergence or overfit, while smaller rank may underfit alignment.Source: G. Ablation Study · p. 11, Table VIII · p. 11

V. CONCLUSION

  • PEARL combines LWDDI for domain-invariant features and LWSD for detection-critical alignment, selecting a pretrained model by feature similarity for cooperative detection.The conclusion describes anonymous real-time assignment and lightweight offline adaptation for new agent types.Source: §V · p. 11

APPENDIX A LIDAR ENCODERS DETAILS

The appendix specifies the LiDAR encoders and training setup, then details PEARL’s parameter footprint and anonymous interpreter selection on OPV2V. It reports selection gains for both familiar and unseen joining agents.

APPENDIX A LIDAR ENCODERS DETAILS

  • The encoder configurations vary voxel resolution and 3D CNN/2D BEV backbones, then compress BEV features to the ego size of 64 × 64 × 128.The appendix also specifies dataset splits, AP@0.5/AP@0.7 evaluation, runtime measurement, and peak inference-memory reporting.Source: Appendix A · p. 13, Appendix B · p. 13

APPENDIX B ADDITIONAL TRAINING AND EVALUATION DETAILS

  • Stage 1 optimizes attention, compressors, extractors, foreground estimators, prompts, and domain classifiers while freezing encoders and the ego stack; Stage 2 updates only the new agent’s adaptation modules.Reported results use validation AP@0.5 checkpoint selection and 28 separately trained configurations.Source: Appendix B · p. 13

APPENDIX C COMPARISON OF TRAINABLE PARAMETER COUNTS

  • Across both training stages, PEARL uses fewer trainable parameters than PolyInter in every scenario except EfficientNet while achieving better detection performance.Stage 1 reduces trainable parameters by 1.1 M–4.4 M across heterogeneous bases; Stage 2 has comparable counts for pp4 and sd1 and higher counts in other cases.Source: Appendix C · p. 13

APPENDIX D DETAILED REAL-TIME IMPLEMENTATION RESULTS

  • Runtime selection applies each pretrained LWDDI to the joining agent’s intermediate BEV features and assigns the interpreter with the highest cosine similarity to the ego’s domain-invariant features, without metadata.The process uses domain-invariant features because other agents’ domain-specific features would require their encoders, exposing sensor/model details and adding computation.Source: Appendix D · pp. 13–14, A. OPV2V · p. 14

A. OPV2V

  • On OPV2V, pretrained interpreters gain up to +24.5% AP@0.5 and +31.2% AP@0.7 for pp4; selection also improves unseen pp6 by +4.5% AP@0.5 and +5.5% AP@0.7 over the all-interpreter average.For vn4 and sd1, pretrained interpreters also outperform random selection on average: +8.1% AP@0.5 and +13.3% AP@0.7 for vn4, and +10.0% AP@0.5 and +18.9% AP@0.7 for sd1.Source: A. OPV2V · p. 14

B. V2XSet

On V2XSet and DAIR-V2X, cosine-similarity selection improves detection over random interpreter selection, including for less-matched agents. The appendix also reports selection latency and a cross-dataset case where the best interpreter differs from the joining agent’s nominal type.

B. V2XSet

  • On V2XSet, matched interpreters improve over random selection by +8.0% AP@0.5 and +12.8% AP@0.7 for pp4, and +7.1% AP@0.5 and +12.9% AP@0.7 for vn4; unseen pp6 gains +2.5% AP@0.5 and +2.1% AP@0.7.Similarity-based selection reliably identifies suitable interpreters while maintaining small selection overhead.Source: B. V2XSet · pp. 14–15

C. DAIR-V2X

  • On DAIR-V2X, matched interpreters improve over random selection by +2.8% AP@0.5 and +1.5% AP@0.7 for pp4, and +2.9% AP@0.5 and +1.7% AP@0.7 for vn4; unseen pp6 gains +0.3% AP@0.5 and +0.6% AP@0.7.The pp6 detection gain is modest, though its similarity score follows the trend on OPV2V and V2XSet.Source: C. DAIR-V2X · p. 15

E. Detailed Cross-Dataset Interpreter Selection

  • Under cross-dataset shift, the highest-similarity sd1 Stage-2 interpreter is also best, reaching 89.8% AP@0.5 and 63.4% AP@0.7, despite the joining agent using pp4.The best available interpreter therefore does not necessarily match the joining agent’s nominal model metadata.Source: E. Detailed Cross-Dataset Interpreter… · p. 16, Table XXIII · p. 20

D. Real-Time Complexity

  • 1.67 ms is the average Pipeline-2 selection time, while the multi-scale extractor takes 47.0 ms per sample, yielding 48.67 ms end-to-end.Selection overhead is on the ego side and requires no additional encoder queries or execution.Source: D. Real-Time Complexity · p. 16, Table XXII · p. 20

APPENDIX E ADDITIONAL ABLATION STUDY

The ablations compare detection performance across ego BEV sizes and detection-head widths, then test LWDDI similarity under alternative extractor, prompt-rank, and feature-resolution choices.

A. Ego BEV Feature and Detection Head Size

  • The default 64 × 64 × 128 ego BEV with a 384-channel head balances detection performance and speed; the larger 64 × 128 × 256 BEV with a 384-channel head raises AP@0.7 by 5.0/2.9% for Stage-1/Stage-2 but adds 4.3 M trainable parameters.The default setting also beats the 256-channel, 64 × 64 × 128 configuration by 0.8% AP@0.5 on average across both stages.Source: A. Ego BEV Feature and Detection Head… · p. 21

B. LWDDI Ablation Study

  • The default LWDDI configuration—multi-scale extraction, PARAFAC prompt rank [8,16,32], and 64×64×128 ego BEV—achieves the highest and most consistent similarity across heterogeneous agents.Removing multi-scale design or sharing an extractor lowers similarity, while larger ego BEV resolution also lowers similarity in this setup; dense 3D prompts score slightly higher for vn6 but use more prompt parameters.Source: B. LWDDI Ablation Study · p. 21
Loading 2609.27123v1…