Aligning Human Motion Generation with Human Perceptions: Additional Visuals

Table of Contents

  1. Annotation Results from MotionPercept Dataset
  2. Results on MotionPercept Test Set
  3. User Study Results of Different Finetuning Steps
  4. Results on OOD AMASS Motions

Part 1: Annotation Results from MotionPercept Dataset

We present videos of four motions along with their original text descriptions and annotator evaluation results. These examples provide insight into how the MotionPercept dataset is annotated and evaluated, as described in Section 3.2 in the paper.

mdma-100-00-04-ABCD.mp4:

Text description: Warm-up

Annnotation result: C

mdma-132-02-02-ABCD.mp4:

Text description: Run

Annnotation result: D

mdma-191-11-03-ABCD.mp4:

Text description: Throw

Annnotation result: C

mdmu-130-35-00-ABCD.mp4:

Text description: Rope-skipping

Annnotation result: A

mdmu-s01-12-01-ABCD.mp4:

Text description:Punching

Annnotation result: C

mdmu-s03-06-01-ABCD.mp4:

Text description: Left-stretching

Annnotation result: C

mdma-100-03-02-ABCD.mp4:

Text description: Jump

Annnotation result: B

mdma-135-07-03-ABCD.mp4:

Text description: Eat

Annnotation result: D

mdma-194-11-01-ABCD.mp4:

Text description: Throw

Annnotation result: D

mdma-143-08-00-ABCD.mp4:

Text description: Turn-steering-wheel

Annnotation result: D

Part 2: Results on MotionPercept Test Set

This section shows videos of four motions, annotation results, MotionCritic scores, and the values of additional metrics as described in Section 3.3 in the paper.

mdma-107-04-02-ABCD.mp4:
Metrics A B C D
Critic score ↑ -0.7886 -4.8514 0.4993 -2.0689
PFC ↓ 2.569e-5 6.449e-5 1.232e-5 3.187e-5
PoseNDF ↓ 0.00039 0.00060 0.00027 0.00080

Text description: Drink

Annnotation result: C

mdmu-s64-11-01-ABCD.mp4:
Metrics A B C D
Critic score ↑ -9.1875 -8.0678 4.8846 1.4563
PFC ↓ 3.601e-6 3.043e-4 4.813e-5 1.025e-4
PoseNDF ↓ 0.00058 0.00051 0.00057 0.00055

Text description: Pulling-chest-expanders

Annnotation result: C

mdma-s84-01-00-ABCD.mp4:
Metrics A B C D
Critic score ↑ -4.4318 -3.4696 -6.5542 2.2963
PFC ↓ 1.045e-3 2.793e-4 2.975e-4 2.958e-4
PoseNDF ↓ 0.00031 0.00054 0.00049 0.00043

Text description: Walk

Annnotation result: D

mdma-131-10-00-ABCD.mp4:
Metrics A B C D
Critic score ↑ -2.1850 1.5277 -0.6620 3.8842
PFC ↓ 3.111e-5 2.580e-5 1.903e-4 1.165e-5
PoseNDF 0.00064 0.00065 0.00037 0.00073

Text description: Boxing

Annnotation result: D

mdmu-112-09-00-ABCD.mp4:
Metrics A B C D
Critic score ↑ 0.8814 -5.0319 0.0365 1.3240
PFC ↓ 3.532e-5 3.245e-4 3.503e-5 2.693e-4
PoseNDF ↓ 0.00039 0.00042 0.00040 0.00041

Text description: Rotation-clapping

Annnotation result: D

mdma-s52-07-01-ABCD.mp4:
Metrics A B C D
Critic score ↑ 0.6195 -2.6155 -2.6123 -2.1859
PFC ↓ 1.292e-6 8.174e-6 8.254e-6 5.938e-6
PoseNDF ↓ 0.00062 0.00107 0.00107 0.00022

Text description: Eat

Annnotation result: A

mdma-s51-01-01-ABCD.mp4:
Metrics A B C D
Critic score ↑ -4.3979 -4.3952 0.2009 -4.4393
PFC ↓ 1.056e-3 1.044e-3 3.007e-4 1.059e-3
PoseNDF ↓ 0.00062 0.00107 0.00107 0.00022

Text description: Eat

Annnotation result: C

mdmu-100-26-01-ABCD.mp4:
Metrics A B C D
Critic score ↑ -0.9983 -12.1793 -3.6499 -10.8046
PFC ↓ 7.699e-5 5.269e-4 5.075e-4 6.801e-4
PoseNDF ↓ 0.00042 0.00052 0.00048 0.00049

Text description: Dumbbell-side-bend

Annnotation result: A

mdmu-s16-16-01-ABCD.mp4:
Metrics A B C D
Critic score ↑ -7.3115 -4.5203 -3.3664 -0.6248
PFC ↓ 4.820e-5 6.480e-5 7.905e-5 2.555e-5
PoseNDF ↓ 0.00091 0.00077 0.00056 0.00062

Text description: Elbow-circling

Annnotation result: D

mdmu-s22-33-00-ABCD.mp4:
Metrics A B C D
Critic score ↑ -2.8390 0.8926 -4.1106 -2.1317
PFC ↓ 1.047e-3 8.792e-4 2.307e-3 8.646e-4
PoseNDF ↓ 0.00049 0.00059 0.00040 0.00042

Text description: Alternate-knee-lifting

Annnotation result: B

Part 3: User Study Results of Different Finetuning Steps

We present video pairs comparing different stages of finetuning, showing the scores provided by MotionCritic and the user preferences collected in the study, as discussed in Section 4 in the paper. This helps confirm the finetuning process through user validation.

step_0 (left) VS step_200 (right) :

Critic score: -13.2

Critic score: -3.79

User preference: step_200

step_400 (left) VS step_600 (right) :

Critic score: 0.995

Critic score: 11.6

User preference: step_600

step_200 (left) VS step_400 (right) :

Critic score: -3.58

Critic score: 1.17

User preference: step_400

step_0 (left) VS step_200 (right) :

Critic score: -3.23

Critic score: 1.49

User preference: step_200

step_200 (left) VS step_400 (right) :

Critic score: -1.05

Critic score: 1.29

User preference: step_400

step_200 (left) VS step_400 (right) :

Critic score: -3.79

Critic score: -0.727

User preference: step_400

step_0 (left) VS step_100 (right) :

Critic score: -2.49

Critic score: -1.79

User preference: step_100

step_100 (left) VS step_300 (right) :

Critic score: -0.761

Critic score: 2.93

User preference: step_300

step_300 (left) VS step_500 (right) :

Critic score: 0.605

Critic score: 1.63

User preference: step_500

step_500 (left) VS step_700 (right) :

Critic score: -0.387

Critic score: 3.45

User preference: step_700

step_0 (left) VS step_200 (right) :

Critic score: -2.70

Critic score: -0.968

User preference: step_200

step_200 (left) VS step_400 (right) :

Critic score: -0.968

Critic score: 0.731

User preference: step_400

step_600 (left) VS step_800 (right) :

Critic score: 0.00632

Critic score: 1.47

User preference: step_800

step_200 (left) VS step_800 (right) :

Critic score: -3.58

Critic score: 4.37

User preference: step_800

step_0 (left) VS step_200 (right) :

Critic score: -4.31

Critic score: -1.05

User preference: step_200

step_0 (left) VS step_200 (right) :

Critic score: 1.294

Critic score: 1.989

User preference: similar

step_600 (left) VS step_800 (right) :

Critic score: 1.989

Critic score: 3.045

User preference: step_800

Part 4: Results on OOD AMASS Motions

To demonstrate generalization capablities of MotionCritic, we provide videos from the AMASS ground truth dataset, along with the critic scores assigned, as discussed in Appendix Section B.3. This highlights the performance of MotionCritic even on out-of-distribution motions.

HumanML3D from AMASS, motion id 002677

Critic score: -16.789180755615234

HumanML3D from AMASS, motion id 002603

Critic score: -12.749826431274414

HumanML3D from AMASS, motion id 001096

Critic score: -12.77638053894043

HumanML3D from AMASS, motion id 000649

Critic score: -12.838809967041016

HumanML3D from AMASS, motion id 002014

Critic score: -13.915456771850586

HumanML3D from AMASS, motion id 000416

Critic score: -14.838180541992188

HumanML3D from AMASS, motion id 001444

Critic score: -15.517463684082031

HumanML3D from AMASS, motion id 000423

Critic score: 10.18864917755127

HumanML3D from AMASS, motion id 000905

Critic score: 10.4847993850708

HumanML3D from AMASS, motion id 000766

Critic score: 10.672083854675293