Rejoining the Past, Piece by Piece with physics-driven deep learning
Welcome to WisePanda's interactive interface!
Here you can explore our system for rejoining fragmented bamboo slips.
Please follow the instructions below to begin experimenting with restoration and reassembly tasks.
You will also find more information about the WisePanda project, links to related articles,
and examples showcasing WisePanda in action.
Bamboo slips were a key medium for recording philosophy, law, and daily life in ancient East Asia. Their
durability has preserved valuable content for millennia, offering deep insights into early civilizations.
However, many slips have been fragmented into thousands of pieces, posing a major challenge to
reconstruction. We present WisePanda, the first physics-driven deep learning framework for rejoining
fragmented bamboo slips. Designed to support archaeologists, WisePanda features an interpretable pipeline
trained on fracture-based synthetic data and provides Top-k predictions to assist expert analysis.
Restoration of fragmented bamboo slips.
Compared to the leading curve matching and modern generative methods, WisePanda increases Top-50 matching
accuracy from 34% and 38% to 52% with statistical significance among more than one
thousand candidate fragments, accelerating archaeologists's efficiency in rejoining fragmented bamboo
slips.
Top-K Accuracy Results
Dataset
Method
Top-1 (%)
Top-5 (%)
Top-10 (%)
Top-20 (%)
Top-50 (%)
Top-100 (%)
Bamboo236
DTW
8.05±0.42
22.88±0.43
31.36±0.43
48.73±1.47
74.72±2.00
94.63±0.89
FMM
8.90±0.85
18.64±0.00
26.41±0.65
37.71±1.53
65.25±1.70
94.35±0.24
SIS
10.45±0.24
21.47±0.98
29.38±2.34
44.07±4.43
71.47±2.45
94.07±0.00
Event-DTW
7.63±0.00
24.01±0.65
31.35±0.73
48.45±1.07
75.42±1.12
94.63±0.25
Drop-DTW
8.34±1.22
23.02±0.65
31.78±0.42
49.29±0.25
75.56±0.65
94.92±0.00
Modified COW
8.05±0.85
24.44±0.25
38.42±1.71
57.06±1.22
85.03±0.65
98.59±0.65
Two-stage DTW
7.63±0.00
23.45±0.49
30.65±0.24
49.44±0.25
74.58±0.00
95.34±0.00
GAN
4.52±1.07
17.94±1.07
32.06±3.66
54.52±2.45
81.36±0.85
96.61±0.73
seriesGAN
4.38±2.69
19.49±4.04
32.49±4.11
52.97±3.83
83.62±1.07
96.75±0.65
DDPM
3.39±0.42
10.03±0.88
19.63±2.55
34.75±0.85
70.48±5.52
96.05±0.98
Diffusion-TS
3.39±1.47
14.83±1.27
27.26±2.82
45.91±2.82
80.37±3.55
98.73±0.85
WisePanda
13.42±1.07
37.15±1.76
50.85±1.53
66.81±2.89
91.81±2.01
100.00±0.00
Bamboo1350
DTW
3.95±0.25
11.72±0.24
14.12±0.25
19.78±0.65
29.80±1.91
37.29±1.12
FMM
5.79±1.36
11.02±0.00
12.85±0.25
15.40±0.49
23.45±1.49
32.20±1.53
SIS
7.34±0.25
14.12±1.29
17.94±1.61
22.04±1.85
32.35±3.45
44.21±4.53
Event-DTW
4.09±0.49
12.85±0.65
15.96±1.07
22.46±1.47
34.04±0.88
42.94±2.00
Drop-DTW
4.24±0.43
11.16±0.24
14.41±0.43
19.92±0.00
29.66±0.00
37.29±0.74
Modified COW
2.97±0.85
9.18±0.65
13.70±1.48
19.35±1.71
28.96±1.71
41.53±2.55
Two-stage DTW
4.10±0.25
12.29±0.00
14.41±0.00
19.78±0.65
29.38±0.24
36.86±0.43
GAN
0.56±0.65
4.24±0.85
7.77±1.71
14.69±3.60
26.27±5.73
41.95±8.35
seriesGAN
1.69±1.47
6.78±0.73
12.43±1.91
21.33±2.18
38.28±3.53
49.15±4.78
DDPM
0.42±0.43
2.26±0.49
3.53±1.07
5.23±1.07
12.71±1.70
21.47±2.59
Diffusion-TS
1.55±1.30
4.52±1.36
8.76±2.09
15.54±2.72
27.97±1.12
43.08±2.89
WisePanda
5.37±1.76
16.53±1.70
25.42±1.12
37.57±0.88
52.54±0.85
64.40±2.24
Wood670
DTW
2.60±0.08
6.71±0.37
11.45±1.13
16.89±1.28
23.61±1.92
29.06±1.18
FMM
2.19±0.09
6.92±0.22
9.35±0.31
13.28±0.00
21.64±0.30
27.41±0.17
SIS
2.34±0.17
7.81±0.17
10.75±0.00
14.92±0.51
23.08±0.60
27.96±0.00
Event-DTW
2.84±0.25
7.06±0.52
10.95±0.08
15.87±0.62
23.38±0.23
28.81±0.26
Drop-DTW
2.39±0.30
6.87±0.30
10.90±0.15
15.42±0.31
23.33±0.17
28.76±0.23
Modified COW
0.95±0.22
2.89±0.45
5.02±0.62
7.86±1.12
14.53±1.45
22.59±1.25
Two-stage DTW
2.74±0.09
6.67±0.31
10.70±0.17
15.87±0.23
23.13±0.15
28.66±0.15
GAN
0.60±0.00
2.24±0.26
4.28±0.31
6.96±0.34
14.48±0.83
24.53±0.90
seriesGAN
0.60±0.15
2.94±0.37
5.07±0.91
8.41±1.25
17.21±0.75
26.97±0.56
DDPM
0.40±0.09
1.29±0.23
2.59±0.43
4.73±0.70
10.45±1.42
18.16±0.99
Diffusion-TS
0.55±0.23
2.14±0.61
4.33±0.65
7.86±1.13
15.92±1.04
25.17±1.53
WisePanda
1.89±0.19
10.14±0.21
17.25±1.56
24.31±2.70
31.24±1.25
35.61±1.65
WisePanda[Wood3833]
1.11±0.29
2.82±0.19
6.67±1.39
9.48±1.78
16.72±1.27
18.38±1.08
This work demonstrates how models like WisePanda can enhance collaboration between AI and historians,
fundamentally transforming how we study and interpret the rich and complex history of China.
WisePanda was conceived and researched by Jinchi Zhu and Yongchao Xu. Hailong Lei, Xiaoguang Wang, Jialiang
Lu and Jing Li provided archaeological data and domain knowledge. Jinchi Zhu and Zhou Zhao performed
experiments, analyzed data, wrote the code and manuscript. Jinchi Zhu, Qianqian Tang and Jiachen Shen
contributed to the literature review and software implementation. Gui-Song Xia participated in methodology
discussions , data analysis, and manuscript organization. Bo Du participated in methodology discussions, data
analysis, and
manuscript editing. Yongchao Xu participated in methodology discussions, data analysis, experimental
designs, manuscript
organization and revision.
Welcome to WisePanda's interactive interface!
