TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos

Patt Phurtivilai; Zhiyang Dou; Yifan Wu; Kinfung Chu; Yuan Liu; Lei Yang; Wenping Wang; Taku Komura

cs.CVSep 29, 2026

TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos

Authors: Patt Phurtivilai, Zhiyang Dou, Yifan Wu, Kinfung Chu, Yuan Liu, Lei Yang, Wenping Wang, Taku Komura

Organizations: The University of Hong Kong, Hong Kong, China · Centre for Transformative Garment Production · Hong Kong University of Science and Technology, Hong Kong, China · Texas A&M University, College Station, TX, US

Abstract

Quantifying collective fish behavior requires accurate trajectories, yet multi-view 3D tracking remains challenging due to frequent occlusions, visually similar individuals, and the long-standing scarcity of identity annotations. We present TrackFish3D, a geometry-driven self-supervised framework for dense multi-camera 3D tracking of schooling fish. Instead of relying on appearance-based re-identification or manually annotated identities, TrackFish3D turns calibrated multi-view geometry into supervision: triangulation and reprojection consistency provide pseudo-associations, while a geometric encoder and global association transformer learn all-to-all cross-view correspondence within each frame. To make these associations identity-aware, TrackFish3D introduces a self-supervised contrastive objective that separates co-visible individuals in the embedding space, together with a temporal predictor that preserves identities and bridges short occlusions across frames. The resulting model is trained once on unlabeled footage and applied directly to unseen test videos, requiring no cross-view identity labels, temporal annotations, 3D ground truth, appearance features, or test-time optimization. On our benchmark, TrackFish3D improves 3D Multi-Object Tracking Accuracy from 87.7% for the strongest baseline to 95.8%. On the 3D-ZeF zebrafish benchmark, it achieves 81.1% MOTA, compared with 77.4% for the best geometric baseline. TrackFish3D also generalizes beyond fish, achieving strong results on real-world bird tracking.

Figures & tables

			No labels / GT required
	Temp.	Spat.	No XV-ID	No T-ID	No 3D GT	No Vis.
SVMOT	✓	✗	✓	✗	✓	✗
Supervised MVA	✗	✓	✗	✓	✓	✗
Self-supervised MVA	✗	✓	✓	✓	✓	✗
MVMOT	✓	✓	✗	✗	✓	✗
TrackFish3D	✓	✓	✓	✓	✓	✓

Tab. 1: Method families for dense 3D tracking. Temp./Spat. : temporal/spatial association. The central columns indicate whether a family can operate without cross-view identity labels (No XV-ID), temporal identity labels (No T-ID), or 3D ground-truth trajectories (No 3D GT). No Vis. : no visual appearance features.

Fig. 2: Overview of the TrackFish3D framework. Synchronised multi-view video is processed by (A) a Geometric Encoder and (B) an Association Transformer that predicts cross-view pairwise scores, supervised by (C) Geometric Pseudo-Supervision from triangulation consistency through (D) a Self-Supervised Association Loss . Detections are fused via (E) Multi-View Grouping & Validation to produce 3D hypotheses, whose embeddings are separated by (F) a Self-Supervised Contrastive Loss . Across frames, (G) Temporal Linking matches identities and (H) a Self-Supervised Temporal Loss enforces cross-frame consistency, enabling stable 3D trajectory recovery at inference.

Fig. 6: Qualitative 3D trajectory comparison on SynFish test scenes. Columns: ground truth, TrackFish3D, and representative baselines. Colour encodes fish identity; trajectory breaks indicate identity switches or fragmentations.

Method	MOTA $\uparrow$	MT% $\uparrow$	ML% $\downarrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
Hungarian + nearest	77.4%	90.0%	0.0%	6.0	73.0	42.5
Greedy + nearest	76.9%	90.0%	0.0%	10.0	76.0	38.7
3D-SORT	75.6%	90.0%	0.0%	13.5	65.5	43.4
Self-MVA [ 12 ]	50.8%	60.0%	0.0%	21.5	73.5	48.5
ASNet [ 7 ]	0.3%	0.0%	0.0%	72.5	175.0	10.6
SambaMOTR [ 54 ]	$-$ 26.4%	0.0%	100.0%	0.5	1.0	6.5

Tab. 3: 3D tracking results on the 3D-ZeF test set, averaged over 2 sequences (zebra2: 5 fish, zebra4: 5 fish). Best result in each column (among methods with MOTA

{>}0

) is bolded .

Fig. 7: Qualitative 3D trajectory comparison on 3D-ZeF test sequences. TrackFish3D maintains identity consistency through real-world occlusions that fragment or swap baseline trajectories. ReST produces no trajectories because its cross-view association fails to match detections correctly.

Ablation 1: loss components
Variant	MOTA $\uparrow$	MT% $\uparrow$	ML% $\downarrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
TrackFish3D (full)	97.7%	98.3%	0.0%	0.5	3.1	283.1
w/o $\mathcal{L}_{\mathrm{temp}}$	86.0%	85.2%	8.0%	2.0	5.2	243.1
w/o $\mathcal{L}_{\mathrm{ctr}}$	36.8%	23.8%	36.4%	3.0	8.0	99.1
w/o both	36.9%	25.6%	34.2%	3.8	9.8	88.8
Ablation 2: visual features

Tab. 4: Ablation studies on SynFish. Top: removing contrastive (

\mathcal{L}_{\mathrm{ctr}}

) and/or temporal (

\mathcal{L}_{\mathrm{temp}}

) losses. Bottom: concatenating DINOv3 [ 56 ] visual features with geometric features.

Variant	MOTA $\uparrow$	MT% $\uparrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
TrackFish3D (SAM3)	95.8%	95.8%	0.0	6.5	266.3
MLP-only + all losses	78.4%	64.2%	2.0	16.3	172.9
SAM3 + Hungarian	83.5%	71.9%	7.5	12.8	183.8

Tab. 5: Controlled Association Transformer ablation on SynFish.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Module	Parameters	%
Geometric Feature Encoder $\phi$	9,216	1.4%
Association Transformer $\psi$	579,457	88.5%
Temporal Predictor $\rho$	66,432	10.1%
Total (trainable)	$\approx$ 655K	100%

Appendix

Tab. 6: TrackFish3D parameter budget (geometry-only mode, no visual backbone).

Category	Hyperparameter	Value
Optimiser	Algorithm	AdamW
	Learning rate	$1\times 10^{-4}$
	Weight decay	$1\times 10^{-4}$
	Gradient clipping (max norm)	5.0
Loss weights	$\lambda_{\mathrm{asso}}$ (association)	1.0
	$\lambda_{\mathrm{ctr}}$ (contrastive)	0.5

Appendix

Tab. 7: Complete TrackFish3D training hyperparameters.

Split	Scene	#Fish	#Frames	#Cams	Resolution	OcclLevel	%Overlap
SynFish (train)	train1	4	359	3	1920 $\times$ 1080	Low	0.2%
SynFish (train)	train2	6	349	3	1920 $\times$ 1080	Mild	10.2%
SynFish (train)	train3	8	362	3	1920 $\times$ 1080	Moderate	81.1%
SynFish (train)	train4	10	359	3	1920 $\times$ 1080	Moderate	75.1%
SynFish (train)	train5	14	359	3	1920 $\times$ 1080	Heavy	91.9%
SynFish (train)	train6	14	362	3	1920 $\times$ 1080	Mild	66.9%

Appendix

Tab. 8: Dataset summary and difficulty statistics. Occlusion level is derived from the OccScore, which is computed per scene by taking, for each (frame, camera) pair, the maximum pairwise 2D bounding-box IoU among all tracked objects, then averaging these maxima across all frames and camera views. %Overlap is the fraction of (frame, camera) instances with any overlap (IoU

>

0.01).

Rig	Train scenes	Test scenes
0	—	test1, test2
1	train1	—
2	train2	—
3	train4	—
4	train5, train6	—
5	train3	test3

Appendix

Tab. 9: SynFish scene-to-rig mapping. Scenes sharing a rig use identical camera placements (same

R

and

\mathbf{t}

for all 3 cameras). Rig 0 appears only in the test set.

Rig	Cam	$R$ (row-major)			$\mathbf{t}$
0	0	$-$ 0.6129	0.7902	$-$ 0.0000	7.48	18.22	43.67
		0.1589	0.1232	$-$ 0.9796
		$-$ 0.7740	$-$ 0.6004	$-$ 0.2011
	1	$-$ 0.9866	$-$ 0.1633	0.0000	$-$ 1.35	13.66	73.22
		$-$ 0.1450	0.8759	$-$ 0.4602
		0.0752	$-$ 0.4540	$-$ 0.8878

Appendix

Tab. 10: SynFish camera rig extrinsics. Each rig defines 3 cameras (0, 1, 2) with rotation

R

(shown row-by-row) and translation

\mathbf{t}

. All cameras share the intrinsic

K_{\mathrm{syn}}

above.

Sequence	Cam	$R$ (row-major)			$\mathbf{t}$
zebra01 (train)	0	0.9999	$-$ 0.0143	0.0054	$-$ 14.68	$-$ 16.34	31.24
		0.0139	0.9974	0.0708
		$-$ 0.0064	$-$ 0.0707	0.9975
	1	0.9999	$-$ 0.0077	$-$ 0.0067	$-$ 14.63	$-$ 9.80	48.50
		0.0079	0.1591	0.9872
		$-$ 0.0065	$-$ 0.9872	0.1591

Appendix

Tab. 11: 3D-ZeF camera extrinsics per sequence. All sequences share the intrinsics

K_{0}

K_{1}

above.

Method	test1	test2	test3	test4	test5	test6	Avg	test1	test2	test3	test4	test5	test6	Avg	test1	test2	test3	test4	test5	test6	Avg
	MOTA (%) $\uparrow$							MT (%) $\uparrow$							ML (%) $\downarrow$
TrackFish3D (SAM3)	99.6	99.9	96.2	86.2	94.8	98.0	95.8	100	100	100	81.2	93.8	100	95.8	0	0	0	0	0	0	0.0
TrackFish3D (YOLO26+SAM2)	100	100	97.7	86.3	96.7	99.1	96.6	100	100	100	81.2	93.8	100	95.8	0	0	0	0	0	0	0.0
SAM3+Hung.	100	100	93.8	55.2	75.1	77.1	83.5	100	100	100	31.2	43.8	56.2	71.9	0	0	0	0	0	0	0.0
SAM3+Greedy	100	100	93.8	55.2	75.1	77.1	83.5	100	100	100	31.2	43.8	56.2	71.9	0	0	0	0	0	0	0.0
YOLO26+SAM2+Hung.	99.7	99.8	95.9	53.7	69.0	75.0	82.2	100	100	100	25.0	43.8	50.0	69.8	0	0	0	0	0	0	0.0

Appendix

Tab. 12: Per-scene accuracy metrics on the SynFish test set (6 held-out scenes). Scene names correspond to Table 8 : test1 (4 fish), test2 (5 fish), test3 (8 fish), test4–test6 (16 fish each).

Method	test1	test2	test3	test4	test5	test6	Avg	test1	test2	test3	test4	test5	test6	Avg	test1	test2	test3	test4	test5	test6	Avg
	IDSw $\downarrow$							Frag $\downarrow$							MTBF $\uparrow$
TrackFish3D (SAM3)	0	0	0	0	0	0	0.0	0	0	2	18	9	10	6.5	359.0	358.6	276.4	158.6	224.7	220.5	266.3
TrackFish3D (YOLO26+SAM2)	0	0	0	0	0	1	0.2	0	0	0	11	5	6	3.7	359.0	359.0	350.6	198.5	271.2	240.5	296.5
SAM3+Hung.	0	0	5	12	17	11	7.5	0	0	6	27	25	19	12.8	359.0	359.0	142.4	66.0	77.8	98.3	183.8
SAM3+Greedy	0	0	5	12	17	11	7.5	0	0	6	27	25	19	12.8	359.0	359.0	142.4	66.0	77.8	98.3	183.8
YOLO26+SAM2+Hung.	0	0	3	21	13	15	8.7	0	0	7	27	18	21	12.2	358.0	358.4	154.1	54.1	88.2	85.9	183.1

Appendix

Tab. 13: Per-scene identity stability metrics on the SynFish test set (6 held-out scenes). Same scene abbreviations as Table 12 .

Method	z2	z4	Avg	z2	z4	Avg	z2	z4	Avg	z2	z4	Avg	z2	z4	Avg	z2	z4	Avg
	MOTA% $\uparrow$			MT% $\uparrow$			ML% $\downarrow$			IDSw $\downarrow$			Frag $\downarrow$			MTBF $\uparrow$
TrackFish3D	82.7	79.4	81.1	100	80	90.0	0	0	0.0	2	6	4.0	24	37	30.5	103.9	83.3	93.6
Hung.+NN	77.1	77.7	77.4	100	80	90.0	0	0	0.0	6	6	6.0	60	86	73.0	43.8	41.2	42.5
Greedy+NN	76.3	77.4	76.9	100	80	90.0	0	0	0.0	8	12	10.0	64	88	76.0	40.3	37.1	38.7
Self-MVA	84.7	16.8	50.8	100	20	60.0	0	0	0.0	2	41	21.5	33	114	73.5	81.2	15.8	48.5
ASNet	13.9	$-$ 13.4	0.3	0	0	0.0	0	0	0.0	55	90	72.5	152	198	175.0	11.5	9.7	10.6

Appendix

Tab. 14: Per-scene 3D tracking results on the 3D-ZeF test set (zebra2: 5 fish, zebra4: 5 fish). ReST fails to produce any valid cross-camera associations on zebrafish sequences.

$\alpha$	$\beta$	MOTA $\uparrow$	MT% $\uparrow$	ML% $\downarrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
1.0	0.0	95.8%	95.8%	0.0%	0.0	6.5	266.3
0.8	0.2	95.8%	95.8%	0.0%	0.0	6.5	266.3
0.6	0.4	95.8%	95.8%	0.0%	0.0	6.5	266.3
0.4	0.6	95.8%	95.8%	0.0%	0.0	6.5	266.3
0.2	0.8	95.8%	95.8%	0.0%	0.2	6.8	262.5
0.0	1.0	95.8%	95.8%	0.0%	0.3	7.5	257.3

Appendix

Tab. 15: Ablation of temporal linking weights

(\alpha,\beta)

on the SynFish test set (6 scenes, averaged). The cost is robust for

\alpha\geq 0.4

; removing the distance component (

\alpha{=}0

) degrades identity stability. The selected setting is bolded .

$\sigma_{R}$ (°)	$\sigma_{t}$	MOTA $\uparrow$	MT% $\uparrow$	ML% $\downarrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
0	0	95.8%	95.8%	0.0%	0.0	6.5	266.3
0.05	0.05	94.1%	95.8%	0.0%	0.0	5.0	274.0
0.1	0.1	92.1%	93.8%	0.0%	0.0	9.2	246.6
0.2	0.2	81.3%	82.3%	1.0%	0.2	20.3	206.0

Appendix

Tab. 16: Robustness of TrackFish3D (SAM 3) to camera extrinsic perturbation on the SynFish test set (6 scenes, averaged).

\sigma_{R}

is the rotation noise standard deviation (degrees) and

\sigma_{t}

is the translation noise standard deviation (world units). The unperturbed baseline (

\sigma{=}0

) is shown in the first row.

$\sigma$	MOTA $\uparrow$	MT% $\uparrow$	ML% $\downarrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
0	95.8%	95.8%	0.0%	0.0	6.5	266.3
0.05	90.3%	95.8%	0.0%	3.8	12.2	175.2
0.10	84.1%	94.8%	0.0%	7.0	15.3	126.0
0.20	72.3%	86.5%	0.0%	11.2	24.8	86.4
0.40	45.4%	40.6%	3.1%	15.3	63.5	37.1

Appendix

Tab. 17: Robustness of TrackFish3D (SAM 3) to detection noise on the SynFish test set (6 scenes, averaged).

\sigma

is the noise level controlling both the false-negative drop rate and the false-positive injection rate. The unperturbed baseline (

\sigma{=}0

) is shown in the first row.

Metric	Value
Parameters	655 K
Training time	17h 36m
Inference (ms / frame)	124.8
Peak GPU memory (GB)	0.02

Appendix

Tab. 18: TrackFish3D runtime and memory profile. Training time is for the full 400-epoch schedule on SynFish (8 training scenes). Inference speed is measured per frame (all cameras), averaged over the SynFish test set.

Family	Methods	Temp	Spat	No Sup	No Vis	How 3D tracks are produced
Det. + geometric MVA	SAM3 / YOLO26+SAM2 + {Hung., Greedy}	–	–	✓	✓	Geometric MVA $\rightarrow$ triangulate $\rightarrow$ NN linking
Det. + geometric MVA + Kalman	SAM3 + 3D-SORT	–	–	✓	✓	Geometric MVA $\rightarrow$ triangulate $\rightarrow$ Kalman tracking
Det. + learned MVA	Self-MVA	–	✓	✓	–	Learned MVA $\rightarrow$ triangulate $\rightarrow$ NN linking
Det. + learned MVA	ASNet	–	✓	–	–	Learned MVA $\rightarrow$ triangulate $\rightarrow$ NN linking
2D MOT $\rightarrow$ 3D	SambaMOTR, MOTIP	✓	–	–	–	2D tracking per view $\rightarrow$ MVA at $t{=}0$ $\rightarrow$ triangulate
Multi-camera MOT	ReST, MCTR	✓	✓	–	–	Cross-view 2D tracking $\rightarrow$ triangulate

Appendix

Tab. 19: Summary of baseline families, capability profiles, and the fair 3D tracking protocol. Temp = learned temporal association; Spat = learned spatial (cross-view) association; No Sup = no identity supervision required; No Vis = no visual appearance features used. All methods receive the same inputs (multi-view detections and calibrated cameras) and use the same post-processing for fairness.

Method	MOTA $\uparrow$	Recall $\uparrow$	Prec. $\uparrow$	IDF1 $\uparrow$	IDSW $\downarrow$	Frag $\downarrow$	MT $\uparrow$	PT $\downarrow$	ML $\downarrow$
Greedy	86.0%	92.6%	95.2%	63.3%	60	36	8	2	0
Hungarian	91.5%	93.9%	97.6%	92.0%	3	6	10	0	0
ASNet [ 7 ]	85.6%	91.6%	95.3%	63.0%	46	40	9	1	0
Self-MVA [ 12 ]	68.7%	78.4%	90.7%	47.8%	50	61	6	4	0
ReST [ 13 ]	88.3%	91.5%	98.5%	58.0%	56	35	8	2	0
MCTR [ 43 ]	28.3%	58.2%	67.7%	35.7%	66	87	1	7	2

Appendix

Tab. 21: Cross-species evaluation on the held-out Pigeon10 sequence seq48 from 3D-POP. TrackFish3D is trained on four real-world pigeon sequences and evaluated on the fifth. Results show that the same geometry-driven formulation transfers beyond fish to another visually similar animal collective.

Dataset	Method	MOTA $\uparrow$	MT% $\uparrow$	ML% $\downarrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
SynFish	SAM3 + Hungarian	83.5%	71.9%	0.0%	7.5	12.8	183.8
SynFish	3D-SORT	87.7%	77.1%	0.0%	4.0	9.0	209.4
SynFish	TrackFish3D (SAM3)	95.8%	95.8%	0.0%	0.0	6.5	266.3
3D-ZeF	Hungarian	77.4%	90.0%	0.0%	6.0	73.0	42.5
3D-ZeF	3D-SORT	75.6%	90.0%	0.0%	13.5	65.5	43.4
3D-ZeF	TrackFish3D	81.1%	90.0%	0.0%	4.0	30.5	93.6

Appendix

Tab. 22: Comparison with the 3D-SORT kinematic baseline.

Method	MOTA $\uparrow$	MT% $\uparrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
TrackFish3D (SAM3)	70.2%	63.3%	8	22	143.7
SAM3 + Hungarian	29.9%	6.7%	40	54	36.0

Appendix

Tab. 23: Single-scene 30-fish SynFish stress test.

Fish	Detections per frame ( $N$ )	Attention entries ( $N^{2}$ )
16	48	2,304
30	90	8,100
32	96	9,216
64	192	36,864
100	300	90,000

Appendix

Tab. 24: Approximate attention scaling with fish count for the three-camera SynFish configuration.

Fish	Precision $\uparrow$	Recall $\uparrow$
4	70.8%	100.0%
16	25.3%	100.0%
30	17.0%	100.0%

Appendix

Tab. 25: Pairwise quality of geometry-derived pseudo-positive associations as density increases.

Training calibration	MOTA $\uparrow$	MT% $\uparrow$	ML% $\downarrow$	IDSw $\downarrow$	Frag $\downarrow$	MTBF $\uparrow$
Clean	81.1%	90.0%	0.0%	4.0	30.5	93.6
Noisy, $\sigma{=}0.1$	66.8%	60.0%	0.0%	13.5	35.5	59.0

Appendix

Tab. 26: 3D-ZeF performance after training with perturbed calibration and evaluating with clean calibration.

Explore similar work

TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking
May 12, 2026Jisu Nam, Jahyeok Koo, Soowon Son +43D Tracking Video Diffusion Models
KAIST AI · Google DeepMind
BEAST3D: Animal behavioral analysis and neural encoding from multi-view video via Gaussian splatting
Jun 1, 2026Yanchen Wang, Lenny Aharon, Wangshu Zhu +73D Representation
Columbia University · Cold Spring Harbor · Stanford University
Fully Distributed Multi-View 3D Tracking in Real-Time
Jun 11, 2026Byron Hernandez, Fangyu Li, Aotian Wu +3Collaborative Fusion
University of Florida, Gainesville, FL, USA · NVIDIA Corporation, Santa Clara, CA, USA

May 12, 2026, cs.CVJ/K move · Enter open · S save

TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking

Jisu Nam, Jahyeok Koo, Soowon Son, Jaewoo Jung+3

KAIST AI · Google DeepMind

3D Tracking Video Diffusion Models Video Latents Discriminative Congruence Transform Prior Knowledge Integration

Open paper AskPDF arXiv

Dense 3D tracking from monocular video is fundamental to dynamic scene understanding. While recent 3D foundation models provide reliable per-frame geometry, recovering object motion in this geometry remains challenging and benefits from strong motion priors learned from real-world videos. Existing 3D trackers either follow iterative paradigms trained from scratch on synthetic data or fine-tune 3D reconstruction models learned from static multi-view images, both lacking real-world motion priors. Pre-trained video diffusion transformers (video DiTs) offer rich spatio-temporal priors from internet-scale videos, making them a promising foundation for 3D tracking. However, their frame-anchored formulation, which generates each frame's content, is fundamentally mismatched with reference-anchored dense 3D tracking, which must follow the same physical points from a reference frame across time. We present TrackCraft3R, the first method to repurpose a video DiT as a feed-forward dense 3D tracker. Given a monocular video and its frame-anchored reconstruction pointmap, TrackCraft3R predicts a reference-anchored tracking pointmap that follows every pixel of the first frame across time in a single forward pass, along with its visibility. We achieve this through two designs: (i) a dual-latent representation that uses per-frame geometry latents and reference-anchored track latents as dense queries, and (ii) temporal RoPE alignment, which specifies the target timestamp of each track latent. Together, these designs convert the per-frame generative paradigm of video DiTs into a reference-anchored tracking formulation with LoRA fine-tuning. TrackCraft3R achieves state-of-the-art performance on standard sparse and dense 3D tracking benchmarks, while running 1.3x faster and using 4.6x less peak memory than the strongest prior method. We further demonstrate robustness to large motions and long videos.

Abstract

Explore similar work

TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking

BEAST3D: Animal behavioral analysis and neural encoding from multi-view video via Gaussian splatting

Fully Distributed Multi-View 3D Tracking in Real-Time