FFBL-Coop: Association-Decoupled Cooperative 3D Multi-Object Tracking

Haoxin Wu; Xiaokai Bai

cs.CVOct 1, 2026

FFBL-Coop: Association-Decoupled Cooperative 3D Multi-Object Tracking

Authors: Haoxin Wu, Xiaokai Bai

Organizations: College of Information Science and Electronic Engineering, Zhejiang University.

Abstract

Cooperative 3D tracking must integrate complementary observations across agents and time while maintaining consistent identities. When evidence integration and identity inheritance share a matching decision, errors arising from cross-view appearance differences and spatial misalignment can compromise both feature fusion and track continuity. We propose FFBL-Coop, a fuse first, bind later framework that separates instance admission from identity management. Confidence-ranked Slot Admission (CSA) allocates cooperative queries to available ego slots using confidence and spatial proximity. Unified Representation Aggregation (URA) uses cooperative semantic features and aligned anchors to guide ego-feature retrieval, refining the augmented query bank within a shared transformer decoder. After refinement, Cooperative-Priority Identity Anchoring (CPIA) combines learned association with persistent mappings to establish accepted identity assignments across frames. A shared codebook reduces transmitted payload while retaining AP and AMOTA close to the uncompressed variant. FFBL-Coop achieves AMOTA/AP of 0.611/0.548 on V2X-Seq and 0.688/0.653 on Griffin-25M. Code will be released.

Figures & tables

Figure 1: (a) Cross-view matching challenges. Top: Geometric discrepancies can make the minimum-cost assignment inconsistent with ground-truth object correspondences. Bottom: The same object has different feature representations and anchor estimates across agent views. (b) System comparison. Detection, tracking, and communication performance on V2X-Seq and Griffin-25M. Annotations show AP/AMOTA in percent and transmission rates in kB/s.

Figure 2: FFBL-Coop framework. Cooperative queries aligned to the ego frame through motion compensation and coordinate transformation. Following fuse first, bind later , CSA admits cooperative evidence into available slots , and URA uses cooperative semantics to guide ego-feature retrieval and refinement. CPIA assigns identities through learned association and ID mappings.

Figure 3: Agent-specific perception stacks. (a) Roadside queries aggregate evidence from constructed BEV features. (b) Aerial queries retrieve evidence from multi-scale FPN features.

Figure 4: Cooperative detection on V2X-Seq. Each row shows vehicle-only, infrastructure-only, and cooperative results in paired image and BEV views. Blue: predictions, red: ground truth.

Figure 5: Comparison with SparseCoop on V2X-Seq. Top: FFBL-Coop, bottom: SparseCoop. Image and BEV pairs include enlarged regions highlighting detection differences.

Method	V2X-Seq			Griffin-25M
Method	AP $\uparrow$	AMOTA $\uparrow$	TC $\downarrow$	AP $\uparrow$	AMOTA $\uparrow$	TC $\downarrow$
No Fusion	16.6	13.0	$0$	37.5	36.5	$0$
Late Fusion	19.6	26.3	$6.60\times 10^{2}$	37.8	37.7	$1.56\times 10^{3}$
Early Fusion	24.3	20.9	$8.19\times 10^{7}$	60.7	67.0	$3.11\times 10^{8}$
V2X-ViT (ECCV 22)	26.8	28.7	$2.56\times 10^{6}$	46.5	50.8	$8.00\times 10^{5}$
UniV2X (AAAI 25)	29.5	23.9	$6.96\times 10^{4}$	41.9	45.6	$5.58\times 10^{4}$

Table 1: System-level comparison on V2X-Seq and Griffin-25M. AP and AMOTA are in percent, TC is transmission cost in B/s. FFBL-Coop uses

K_{\mathrm{code}}=256/512

, respectively. Baselines retain their published architectures and results. † Long-SCOPE reports the

0

–

50\,\mathrm{m}

subset of its long-range evaluation and is included for reference. Bold marks the best results excluding this reference row.

Base	CSA	URA	CPIA	V2X-Seq			Griffin-25M
Base	CSA	URA	CPIA	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$
$\checkmark$				39.9	47.7	123	48.7	48.4	162
$\checkmark$		$\checkmark$		53.2	50.4	116	64.0	55.3	86
$\checkmark$		$\checkmark$	$\checkmark$	54.2	60.1	69	64.1	66.1	57
$\checkmark$	$\checkmark$		$\checkmark$	53.2	59.0	72	60.6	64.9	48
$\checkmark$	$\checkmark$	$\checkmark$		54.3	59.3	78	62.9	67.3	124

Table 2: Module ablations. Disabled CSA, URA, and CPIA use Hungarian allocation, mean aggregation, and slot-ID inheritance, respectively. URA-only uses one Hungarian assignment for admission and identity inheritance. Base retains cooperative input. Bold marks the best displayed results.

Bin	V2X-Seq			Griffin-25M
Bin	$N$	Concat-only	Full	$N$	Concat-only	Full
ego-only	1673	34.9	35.0	178	18.6	17.7
coop-only	2699	62.9	63.1	1569	54.9	55.6
both	1227	39.2	53.5	4959	42.2	72.1

Table 3: AP by camera-projection coverage.

N

: ground-truth count. Concat-only: no in-region fusion.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	AP@0.5	AP@1	AP@2	AP@4	AP $\uparrow$	ATE $\downarrow$	AVE $\downarrow$
V2X-Seq	19.95	49.95	70.05	79.25	54.8	0.504	1.093
Griffin-25M	44.15	66.35	73.05	77.65	65.3	0.291	1.661

Appendix

Table 7: Detection results at centre-distance thresholds of

0.5

1

2

, and

4\,\mathrm{m}

. Overall AP averages the four threshold-specific scores. ATE and AVE are measured in

\mathrm{m}

and

\mathrm{m/s}

, respectively.

Dataset	B/s	TP	FP	FN	IDS	MOTA 0.3	MOTA 0.5	MOTA 0.7	AMOTA $\uparrow$
V2X-Seq	$8.48\times 10^{3}$	3510	551	2000	70	27.7	45.5	52.9	61.1
Griffin-25M	$3.37\times 10^{4}$	4982	669	1790	50	30.9	46.2	60.8	68.8

Appendix

Table 8: Tracking results and transmitted payload rates. TP, FP, FN, and IDS use a

2\,\mathrm{m}

matching threshold. MOTA subscripts denote recall levels. Communication accounting follows Table 1 .

Panel of Figure 6 — Figure 6: Comparison of ego-only, roadside-only, and cooperative detection sequences on V2X-Seq.

Panel of Figure 7 — Figure 7: Cooperative detections on Griffin-25M from both drone and ego-vehicle perspectives.

Dataset / bin	$N$	Base AP	Full AP	$\Delta$	Full R@2m
V2X-Seq / light	2981	64.6	68.8	$+4.2$	95.2
V2X-Seq / occluded	2618	33.3	38.4	$+5.1$	84.0
Griffin-25M / light	6528	35.5	56.4	$+20.9$	70.5
Griffin-25M / occluded	1683	0.9	39.2	$+38.3$	75.0

Appendix

Table 9: Car detection by occlusion severity.

N

counts ground-truth objects. Base uses concat-only input. AP and recall are in percent, and

\Delta

denotes the AP gain in percentage points. R@2m is Full recall at a

2\,\mathrm{m}

matching threshold.

Dataset / range	$N$	Base AP	Full AP	$\Delta$	Full@0.5
V2X-Seq / $0$ – $20\,\mathrm{m}$	1193	58.5	59.8	$+1.3$	29.6
V2X-Seq / $20$ – $40\,\mathrm{m}$	2751	53.4	57.6	$+4.2$	24.2
V2X-Seq / $40$ – $50\,\mathrm{m}$	1655	33.6	40.3	$+6.7$	10.9
Griffin-25M / $0$ – $20\,\mathrm{m}$	3742	31.6	47.3	$+15.7$	33.7
Griffin-25M / $20$ – $40\,\mathrm{m}$	3136	22.3	61.1	$+38.8$	40.3
Griffin-25M / $40$ – $50\,\mathrm{m}$	1333	3.2	27.6	$+24.4$	9.6

Appendix

Table 10: Car detection by ego-frame range.

N

counts ground-truth objects. AP is in percent, and

\Delta

denotes the AP gain over concat-only Base in percentage points. Full@0.5 reports AP at a

0.5\,\mathrm{m}

centre-distance threshold.

Figure 8: Paired ego and roadside views on V2X-Seq. Nearby vehicles obscure objects in the ego view, while the roadside perspective provides complementary observations.

Base	CSA	URA	CPIA	V2X-Seq			Griffin-25M
Base	CSA	URA	CPIA	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$
$\checkmark$				39.9	47.7	123	48.7	48.4	162
$\checkmark$	$\checkmark$			52.6	56.5	73	59.2	63.4	43
$\checkmark$		$\checkmark$		53.2	50.4	116	64.0	55.3	86
$\checkmark$			$\checkmark$	53.0	58.9	75	60.5	63.8	66
$\checkmark$		$\checkmark$	$\checkmark$	54.2	60.1	69	64.1	66.1	57

Appendix

Table 11: Complete module ablations under a shared training protocol. Base uses the stated replacements. Bold marks the best results.AP and AMOTA are reported in percent.

Admission budget				CPIA adoption radius
$K$	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$	$\tau_{\mathrm{id}}$	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$
5	53.6	58.7	37	$0\,\mathrm{m}$	54.7	60.5	76
20	54.6	61.2	51	$1.0\,\mathrm{m}$	54.8	61.1	72
35	54.8	61.1	70	$\mathbf{2.2\,\mathrm{m}}$	54.8	61.1	70
All	54.7	61.0	79	$3.0\,\mathrm{m}$	54.6	59.3	59

Appendix

Table 12: Inference-time sensitivity on V2X-Seq. All removes the budget cap.

\tau_{\mathrm{id}}=0

disables spatial adoption. AP and AMOTA are in percent. Bold marks the defaults.

Admission budget				CPIA adoption radius
$K$	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$	$\tau_{\mathrm{id}}$	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$
10	64.3	67.9	55	$0\,\mathrm{m}$	65.2	68.7	63
25	64.9	68.3	53	$0.75\,\mathrm{m}$	65.3	68.8	59
50	65.3	68.8	50	$\mathbf{1.5\,\mathrm{m}}$	65.3	68.8	50
All	65.3	68.8	51	$3.0\,\mathrm{m}$	65.3	67.9	53

Appendix

Table 13: Inference-time sensitivity on Griffin-25M. All removes the budget cap.

\tau_{\mathrm{id}}=0

disables spatial adoption. AP and AMOTA are in percent. Bold marks the defaults.

Component	Intervention	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$
Reference	Full	54.8	61.1	70
CSA	Early prefix admission	50.5	50.9	67
CSA	$\tau_{\mathrm{CSA}}=3.0\,\mathrm{m}$	54.5	60.0	74
URA	w/o local readout	52.3	54.7	97
URA	Mean-vector feature replacement	51.2	52.9	163
CPIA	Sticky occupancy	54.3	55.3	103

Appendix

Table 14: Detailed ablations on V2X-Seq covering admission, refinement, and CPIA components.

Figure 9: Identity continuity on five consecutive V2X-Seq frames (tokens $010562$ – $010582$ , $2\,\mathrm{Hz}$ ). Top: Full (CPIA). Bottom: CSA+URA with sticky IDs. Different colors denote different track IDs.

Codebook	B/s $\downarrow$	AP $\uparrow$	AMOTA $\uparrow$	IDS $\downarrow$
Uncompressed	$6.71\times 10^{5}$	65.5	68.9	50
$K_{\mathrm{code}}=1024$	$3.37\times 10^{4}$	64.7	68.3	53
$\bm{K_{\mathrm{code}}=512}$	$\bm{3.37\times 10^{4}}$	65.3	68.8	50
$K_{\mathrm{code}}=256$	$3.31\times 10^{4}$	64.3	66.2	68
$K_{\mathrm{code}}=64$	$3.31\times 10^{4}$	64.1	66.2	64
$K_{\mathrm{code}}=1$	$3.31\times 10^{4}$	63.3	65.7	63

Appendix

Table 15: Feature compression on a fixed Griffin-25M Full checkpoint. B/s measures steady-state query transmission, excluding one-time codebook synchronization. AP and AMOTA are in percent. Bold marks the default configuration.

Explore similar work

CDKF-Track: Cluster-aware Data-Driven Kalman Filtering for Cooperative 3D Multi-Object Tracking
Sep 22, 2026Maria Damanaki, Nikos Piperigkos, Alexandros Gkillas +1Multi-Object Tracking 3D Object Detection
Industrial Systems Institute, Athena Research Center, Patras Science Park, Greece · 3Dpt. of Informatics & Telecom., University of Ioannina, Arta, Greece · Avisense.AI, Patras, Greece +1
CoGoal3D: Collaborative 3D Object Detection with 3D-Aware Fusion and Refinement
Jul 21, 2026Zhihao Yang, Zhiyu Xiang, Peng Xu +53D Object Detection V2X
Zhejiang University, Hangzhou, China · Zhejiang Provincial Key Laboratory of Multi-Modal Communication Networks and Intelligent Information Processing, Hangzhou, China · China Jiliang University, Hangzhou, China
Sparse2comm: Towards Robust Cooperative 3D Object Detection
Oct 6, 2026Lei Yang, Boqi Li, Chunmian Lin +63D Object Detection Collaborative Perception
University of Hong Kong, China

Sep 22, 2026, cs.ROJ/K move · Enter open · S save

CDKF-Track: Cluster-aware Data-Driven Kalman Filtering for Cooperative 3D Multi-Object Tracking

Maria Damanaki, Nikos Piperigkos, Alexandros Gkillas, Aris S. Lalos

Industrial Systems Institute, Athena Research Center, Patras Science Park, Greece · 3Dpt. of Informatics & Telecom., University of Ioannina, Arta, Greece · Avisense.AI, Patras, Greece+1

Multi-Object Tracking 3D Object Detection Multi-Sensor Fusion Motion Model Extended Kalman Filter

Open paper AskPDF arXiv

Multi-Object Tracking (MOT) is essential for EdgeAI perception systems, where accurate object localization and reliable identification enable safe decision-making. Singleagent MOT suffers from occlusions, sensor noise, and partial scene understanding in complex real-world scenarios. While multi-agent systems improve robustness by exploiting shared information, they introduce redundant measurements that lead to false data associations, and still struggle to capture nonlinear object dynamics. To address these challenges, we propose CDKFTrack, a Cluster-aware Data-Driven Kalman Filtering framework for Cooperative 3D MOT. The proposed method first fuses multivehicle 3D LiDAR detections through a Graph Laplacian-based formulation. Then, a cluster-aware redundancy reduction scheme groups spatially related detections and selects representative observations to reduce duplicate inputs to the tracker. The resulting detections are processed by a data-driven Kalman filter that learns object motion dynamics from data, reducing dependence on predefined linear motion assumptions. Furthermore, a wavelet-based temporal refinement module leverages the multiresolution decomposition property of wavelets to attenuate shortterm positional fluctuations and improve trajectory continuity. To the best of our knowledge, CDKF-Track is the first framework to jointly address detection-level fusion redundancy and learnable motion modeling in cooperative 3D MOT. Experimental results on the real-world V2V4Real dataset indicate that CDKF-Track achieves up to 27.99% improvements in tracking accuracy over state-of-the-art multi-agent MOT methods.

Abstract

Explore similar work

CDKF-Track: Cluster-aware Data-Driven Kalman Filtering for Cooperative 3D Multi-Object Tracking

CoGoal3D: Collaborative 3D Object Detection with 3D-Aware Fusion and Refinement

Sparse2comm: Towards Robust Cooperative 3D Object Detection