Scriboora: Rethinking Human Pose Forecasting

Daniel Bermuth; Alexander Poeppel; Wolfgang Reif

cs.CVNov 19, 2025

Scriboora: Rethinking Human Pose Forecasting

Authors: Daniel Bermuth, Alexander Poeppel, Wolfgang Reif

Organizations: University of Augsburg, Germany

Abstract

Human pose forecasting predicts future human poses from past observations for applications including action recognition, sports analysis, and human-robot interaction. However, progress in the field is difficult to assess because reported results often rely on heterogeneous preprocessing, inconsistent metric implementations, and partially released artifacts. This paper re-examines human pose forecasting as an evaluation problem by auditing a wide range of recent forecasting methods, identifying several reproducibility issues, and introducing a unified training and evaluation pipeline for absolute pose forecasting. Under this unified protocol, mature cross-domain sequence models adapted from speech recognition provide strong baselines and obtain the lowest errors among the evaluated models. To study deployment-relevant noise, a paired clean/detector-noisy dataset variant is introduced using poses generated by a pose-estimation model. The experiments show that detector-generated poses cause substantial performance degradation compared with clean motion-capture inputs, but that part of this degradation can be recovered by adapting the forecaster on detector-generated pose sequences.

Figures & tables

Figure 1 : Example of an absolute pose forecast of a walking person. The green skeletons visualize the input sequence of the prediction model, the red ones the predicted future poses, and the blue ones the ground-truth labels.

Method	Reproduction outcome	Main issue
MotionMixer [ 5 ]	corrected result lower	averaging/counting bug
STSGCN [ 31 ]	corrected result lower	MPJPE implementation bug
GMFnet [ 30 ]	checkpoint reproduced	training/checkpoint mismatch
SPGSN [ 20 ]	retrained result lower	unresolved training mismatch
HisRepItself [ 38 ]	reproduced by code	–
AuxFormer [ 40 ]	checkpoint reproduced	training/checkpoint mismatch

Table 1 : Summary of reproducibility audit. The table reports the dominant issue observed for each method. “Checkpoint reproduced” means that the released checkpoint matched the reported result, but training from the released code did not reproduce the same performance. “Reproduced by code” means that the model was successfully trained and evaluated using the publicly released code. “Corrected result lower” means that an evaluation issue was identified and the corrected score was worse than the originally reported score. Detailed numerical comparisons and forensic notes are provided in Appendix B .

Method	Type	Size	MPJPE	FPS	FCE	FADE
Repeat last frame	-	-	301	-	-	301
Last delta average	-	-	264	-	-	264
Ridge Regression	-	-	185	-	-	185
TBiFormer ’2023	abs	5.64M	301	293	7	302
PGBIG ’2022	rel	2.65M	279	35	57	287
GMFnet ’2024	rel	11.3M	279	51	39	284

Table 2 : Unified comparison of absolute and relative methods in the task of absolute pose forecasting on Human3.6M . MPJPE and FADE are measured at timestep

1000\,ms

. The error metrics are in millimeters. FPS is tested on a computer with an AMD-9900X and a single Nvidia-RTX4080. A batch-size of

1

was used for testing. Note that the size and speed of some models depends on the number of input and output timesteps (here:

50

in,

25

out).

Method	Size	MPJPE	FPS	FCE	FADE
DeepSpeech ’2014	2.35M	198	653	3	198
QuartzNet ’2020	3.17M	168	1300	2	168
Squeezeformer ’2022	3.31M	149	960	2	149
Conformer ’2020	3.22M	147	1049	2	147
MotionConformer	9.23M	143	929	2	143

Table 3 : Results of converted speech models on Human3.6M .

Method	Size	MPJPE 400 / 1000	FPS	FADE 400 / 1000
Repeat last frame	-	201 / 408	-	201 / 408
Last delta average	-	167 / 398	-	167 / 398
Ridge Regression	-	130 / 296	-	130 / 296
JRTransformer	3.79M	99.0 / 245	610	99.4 / 245
DeformMLP	1.29M	98.3 / 242	462	98.8 / 243
EqMotion	3.37M	90.6 / 217	41	96.1 / 222

Table 4 : Comparison of single person forecasting on CMU-MoCap dataset. The models were trained to predict a maximum time-range of

1000ms

, the forecast error was measured at the timesteps

400ms

and

1000ms

Method	Size	MPJPE 1000/2000/3000	FPS	FADE 1000/2000/3000
Repeat last frame	-	356 / 562 / 706	-	356 / 562 / 706
Last delta average	-	373 / 773 / 1198	-	373 / 773 / 1198
Ridge Regression	-	276 / 507 / 692	-	276 / 507 / 692
DeformMLP	1.29M	241 / 454 / 656	226	242 / 455 / 657
JRTransformer	4.38M	255 / 446 / 602	562	255 / 446 / 602
EMPMP	3.91M	233 / 385 / 489	201	234 / 386 / 490

Table 5 : Comparison of single person forecasting on CMU-MoCap dataset. The models were trained to predict a maximum time-range of

3000ms

, the forecast error was measured at

1000ms

2000ms

and

3000ms

. Again, twice the output duration was used as input duration. The reason for the difference of repeat last frame method at timestep

1000ms

to the table before is the windowing of the dataset. Here, the window is much larger, resulting in fewer samples, because the windows start later and end earlier.

Method	Size	MPJPE	FPS	FCE	FADE
Repeat last frame	-	272	-	-	272
Last delta average	-	318	-	-	318
Ridge Regression	-	277	-	-	277
EqMotion	3.37M	234	16	125	249
EMPMP	314K	227	209	10	228
MotionConformer	10.2M	207	853	2	207

Table 6 : Forecasting close two-person interactions on CHi3D , errors at timestep

1000\,ms

Method	MPJPE
	$40ms$	$400ms$	$1000ms$
EMPMP	4.3	59.7	165
EqMotion	4.0	52.0	156
MotionConformer	4.0	51.6	149

Table 7 : Zero-shot transfer capabilities from a large mixture of datasets to Human3.6M , at the given timesteps.

Method	detector target		ground-truth target
	$400ms$	$1000ms$	$400ms$	$1000ms$
EqMotion	513	842	532	856
EMPMP	100	215	123	229
MotionConformer	94.2	214	118	228

Table 8 : Influence of skeleton joints generated by a 3D-pose-estimation network on model performance, measured as MPJPE in millimeters at the given timesteps. The left section uses the network predictions as inputs and targets (which would be the measurable error in live systems), and the right one the predictions for input and ground-truth labels as targets (which would be the real error).

Method	detector target		ground-truth target
	$400ms$	$1000ms$	$400ms$	$1000ms$
MotionConformer	66.9	159	94.8	177
EMPMP	76.4	179	104	195
EqMotion	68.5	168	96.1	185
MotionConformer	69.1	165	98.3	183

Table 9 : Improvements through pseudo-label training. Finetuning the pretrained model on noisy predictions in the upper part, training all models from scratch in the lower part. Columns are separated the same as in Table 8

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2 : Example of a relative pose forecast of a walking person. The green skeletons visualize the input sequence of the prediction model, and the red ones the predicted future poses. Only the predicted joints are visualized, therefore this person does not have a hip (it is fixed to the same spot).

Method	MPJPE (400/1000ms)	Replicated
Repeat last frame	88.2 / 136.6	-
Ridge Regression	76.5 / 128.2	-
Res. Sup. [ 24 , 38 ] ’2017	88.3 / 136.6	-
IAFormer [ 39 ] ’2024	50.8 / 130.4	-
MotionMixer [ 5 ] † ’2022	59.3 / 111.0	65.4 / 117.9
STSGCN [ 31 ] † ’2021	38.3 / 75.6	67.5 / 117.0

Appendix

Table 10 : Comparing recent approaches for relative pose forecasting on Human3.6M dataset, averaged over all actions. The mean per joint position error (MPJPE) is measured in millimeters at two frames 400 ms and 1000 ms in the future (using the same model for both timesteps). Approaches marked with a † had errors in their evaluation code which were fixed in the replicated column.

Method	Reporting Paper	MPJPE (1000ms / 3000ms)
HisRepItself	MRT [ 36 ]	1.05 / 1.58
	TBiFormer [ 27 ]	134 / 349
	SoMoFormer [ 34 ]	0.50 / 1.42
	JRTransformer [ 42 ]	14.9 / 30.7
MRT	MRT [ 36 ]	0.79 / 1.22
	TBiFormer [ 27 ]	148 / 352

Appendix

Table 11 : Comparing reported results of recent approaches for absolute pose forecasting on the CMU-MoCap dataset proposed by MRT [ 36 ] . Even though the descriptions of the experiments are similar, the reported results differ substantially between papers, showing general replicability problems.

Figure 3 : Measured MPJPE over time for MotionConformer , together with local one-step extrapolation segments used to validate the FADE approximation. Each segment starts at a measured timestep and follows the linear increase assumed by FADE . The small vertical offset between the segment endpoint and the next measured MPJPE value indicates the local approximation error.

Method	Size	MPJPE	FPS	FADE
Conformer	3.22M	147	1049	147
increased model size	8.41M	144	932	144
moved time reduction	3.59M	146	965	146
MotionConformer	9.23M	143	929	143
no spec augmentation		144
no convolutional modules	7.89M	145	1024	145

Appendix

Table 12 : Implemented improvements and ablations of MotionConformer , tested on Human3.6M .

Method	detector target		ground-truth target
	$400ms$	$1000ms$	$400ms$	$1000ms$
EMPMP	82.5	188	108	204
MotionConformer	82.4	185	106	199

Appendix

Table 13 : Improvements through artificial noise pre-training. The left/right columns are separated into prediction/groundtruth targets, as in Table 8 , showing the measurable/real error.

Action-Class	MPJPE	MPJPE	above	MPJPE	Difference
Joint-Type	(clean)	(input)	10cm %	(noisy)	(noisy-clean)
Directions	140	54	5	156	16
Discussion	174	53	5	196	22
Eating	90	52	6	103	13
Greeting	235	96	37	262	27
Phoning	101	57	8	124	23

Appendix

Table 14 : Errors of MotionConformer per action and joint type on the Human3.6M dataset using ground-truth or detector-noisy joints.

Panel of Figure 4 — Figure 4 : Example of improvements through finetuning on Human3.6M dataset. In blue the ground-truth, in orange the prediction. The walking movement was already well continued in (a), but especially the stride distances improved after finetuning. For better visualization some intermediate timesteps are not displayed.

Explore similar work

HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark
Aug 13, 2026Dairu Liu, Zekun Qi, Jiayu Zeng +11Human Motion Motion Imitation
1Nankai University · 3Galbot · 2Tsinghua University +3
LA-Pose: Latent Action Pretraining Meets Pose Estimation
Apr 30, 2026Zhengqing Wang, Saurabh Nair, Prajwal Chidananda +4Wifi-Csi 3D Human Pose Estimation Latent Action Models
1Wayve · 2Simon Fraser University
Long-Term Prediction of Local and Global Human Motion with Occlusion Recovery
Jun 26, 2026Qiaoyue Yang, Sven Heutger, Christopher Niemann +3Human Motion Prediction Human Motion
Neuro-Information Technology, Otto-von-Guericke-University Magdeburg, Magdeburg, 39106, Germany

Aug 13, 2026, cs.ROJ/K move · Enter open · S save

HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark

Dairu Liu, Zekun Qi, Jiayu Zeng, Ruixi Yu+10

1Nankai University · 3Galbot · 2Tsinghua University+3

Human Motion Motion Imitation Motion Human-Annotated Benchmark

Open paper AskPDF arXiv

Humanoid motion tracking is central to teleoperation and whole-body imitation, yet evaluation often disagrees with what people perceive in videos. Kinematic errors average per-frame pose differences but miss the physical artifacts that matter most, particularly unstable support and incorrect contacts such as foot skating and mistimed touch-downs. Meanwhile, widely used test suites are small and lack the diversity needed to stress contact-rich, long-horizon behaviors. We introduce HumanTracker to make humanoid tracking evaluation both perceptually aligned and scalable. The HumanTracker benchmark contains approximately 153 hours of optical motion trajectories from multiple professional performers, organized into four motion families with text labels for fine-grained diagnosis. We further propose HumanScore, a preference-aligned metric trained on 12K motion pairs containing 24K motions. Across representative state-of-the-art trackers, HumanScore better predicts human preferences and reveals contact and stability failures that kinematic metrics often miss.

Abstract

Explore similar work

HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark

LA-Pose: Latent Action Pretraining Meets Pose Estimation

Long-Term Prediction of Local and Global Human Motion with Occlusion Recovery