PAct input correction

Corrected Full-Object Inputs vs Previous Fragmented Proxy Inputs

这页专门核对你指出的问题:PAct 应该吃完整物体的 RGB 条件图,而不是由分散 part mesh 拼出来的碎片图。下面的 corrected 组用完整物体彩色/材质渲染重新跑了 official PAct、旧预处理 Stage1 cache、以及 corrected-render 重新微调的 Stage1 cache,并保留三维 GLB 与开合视频。

GT parts/joints35/32
Official parts/joints12/8
Stage1-old-preprocess parts/joints12/8
Stage1-corrected-render parts/joints12/9

sample

electronics_103863

GT 3/2 Official 3/2 Old Stage1 3/2 Corrected Stage1 3/3 Corrected delta +0/+1

Previous fragmented input

Previous fragmented input

The proxy part-mesh render that caused the suspicious scattered visual condition.

Corrected PAct RGB input

Corrected PAct RGB input

The full-object color/material render actually used for this corrected inference run.

Corrected full-part mask

Corrected full-part mask

Mask labels preserve all visible parts, including small buttons where present.

20 old training views from fragmented/proxy preprocessing
training view 00
00
training view 01
01
training view 02
02
training view 03
03
training view 04
04
training view 05
05
training view 06
06
training view 07
07
training view 08
08
training view 09
09
training view 10
10
training view 11
11
training view 12
12
training view 13
13
training view 14
14
training view 15
15
training view 16
16
training view 17
17
training view 18
18
training view 19
19
20 corrected training views used by the corrected-render Stage1 fine-tune
training view 00
00
training view 01
01
training view 02
02
training view 03
03
training view 04
04
training view 05
05
training view 06
06
training view 07
07
training view 08
08
training view 09
09
training view 10
10
training view 11
11
training view 12
12
training view 13
13
training view 14
14
training view 15
15
training view 16
16
training view 17
17
training view 18
18
training view 19
19

sample

electronics_104011

GT 28/27 Official 5/4 Old Stage1 5/4 Corrected Stage1 5/4 Corrected delta +0/+0

Previous fragmented input

Previous fragmented input

The proxy part-mesh render that caused the suspicious scattered visual condition.

Corrected PAct RGB input

Corrected PAct RGB input

The full-object color/material render actually used for this corrected inference run.

Corrected full-part mask

Corrected full-part mask

Mask labels preserve all visible parts, including small buttons where present.

20 old training views from fragmented/proxy preprocessing
training view 00
00
training view 01
01
training view 02
02
training view 03
03
training view 04
04
training view 05
05
training view 06
06
training view 07
07
training view 08
08
training view 09
09
training view 10
10
training view 11
11
training view 12
12
training view 13
13
training view 14
14
training view 15
15
training view 16
16
training view 17
17
training view 18
18
training view 19
19
20 corrected training views used by the corrected-render Stage1 fine-tune
training view 00
00
training view 01
01
training view 02
02
training view 03
03
training view 04
04
training view 05
05
training view 06
06
training view 07
07
training view 08
08
training view 09
09
training view 10
10
training view 11
11
training view 12
12
training view 13
13
training view 14
14
training view 15
15
training view 16
16
training view 17
17
training view 18
18
training view 19
19

sample

household_items_102201

GT 4/3 Official 4/2 Old Stage1 4/2 Corrected Stage1 4/2 Corrected delta +0/+0

Previous fragmented input

Previous fragmented input

The proxy part-mesh render that caused the suspicious scattered visual condition.

Corrected PAct RGB input

Corrected PAct RGB input

The full-object color/material render actually used for this corrected inference run.

Corrected full-part mask

Corrected full-part mask

Mask labels preserve all visible parts, including small buttons where present.

20 old training views from fragmented/proxy preprocessing
training view 00
00
training view 01
01
training view 02
02
training view 03
03
training view 04
04
training view 05
05
training view 06
06
training view 07
07
training view 08
08
training view 09
09
training view 10
10
training view 11
11
training view 12
12
training view 13
13
training view 14
14
training view 15
15
training view 16
16
training view 17
17
training view 18
18
training view 19
19
20 corrected training views used by the corrected-render Stage1 fine-tune
training view 00
00
training view 01
01
training view 02
02
training view 03
03
training view 04
04
training view 05
05
training view 06
06
training view 07
07
training view 08
08
training view 09
09
training view 10
10
training view 11
11
training view 12
12
training view 13
13
training view 14
14
training view 15
15
training view 16
16
training view 17
17
training view 18
18
training view 19
19