id
stringlengths
23
47
paper_id
stringlengths
10
10
fig_id
stringclasses
29 values
row
stringlengths
1
25
row_raw
stringlengths
1
25
question
stringlengths
46
114
choices
listlengths
2
4
answer
stringclasses
4 values
answer_known
bool
1 class
methods
listlengths
2
4
image_paths
listlengths
2
4
leaf_name
stringclasses
38 values
leaf_path
listlengths
2
2
leaf_definition
stringclasses
1 value
rationale
stringclasses
1 value
explanation
stringlengths
44
368
must_cite_visual
listlengths
0
0
caption
stringlengths
96
958
fig_refs
listlengths
0
0
v5_visual_attr
stringlengths
12
80
v5_claim_raw
stringlengths
44
368
v5_claim_norm
stringlengths
48
372
⌀
v5_taxonomy_axis
stringclasses
6 values
v5_taxonomy_leaf
stringclasses
38 values
v5_taxonomy_conf
float64
0.65
0.95
v5_data_point_key
stringlengths
22
46
paper_0842__fig_4_p6__row_3
paper_0842
fig_4_p6
row_3
row_3
Which method demonstrates better Accuracy of semantic segmentation labels assigned to 3D point cloud regions?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "Support", "COSeg", "Support Mask", "QGE" ]
[ "paper_0842/figures/fig_4_p6/crops/bbox_08_support.png", "paper_0842/figures/fig_4_p6/crops/bbox_13_coseg.png", "paper_0842/figures/fig_4_p6/crops/bbox_09_support_mask.png", "paper_0842/figures/fig_4_p6/crops/bbox_04_qge.png" ]
Attribute Binding
[ "Relation", "Attribute Binding" ]
Qualitative comparison of 3D point cloud semantic segmentation results between COSeg and QGE on 1-way 1-shot tasks with different target categories (floor, chair, table)
[]
Figure 4. Qualitative comparisons between our proposed model COSeg and QGE [29]. Each row, from top to bottom, represents the 1-way 1-shot task with the target category as floor (blue), chair (red), and table (purple), respectively. a substantial 32.96% mIoU drop is observed from 77.16% to 44.20% for the 5-shot task. O...
[]
Accuracy of semantic segmentation labels assigned to 3D point cloud regions
Qualitative comparison of 3D point cloud semantic segmentation results between COSeg and QGE on 1-way 1-shot tasks with different target categories (floor, chair, table)
Qualitative comparison of 3D point cloud semantic segmentation results between image B and QGE on 1-way 1-shot tasks with different target categories (floor, chair, table)
Relation
Attribute Binding
0.8
paper_0842/fig_4_p6::row_3
paper_0269__fig_6_p6__row_3
paper_0269
fig_6_p6
row_3
row_3
Which method demonstrates better highest level of detail in point cloud reconstructions?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "Pix2Vox++", "Ours", "LegoFormer" ]
[ "paper_0269/figures/fig_6_p6/crops/bbox_06_pix2vox.png", "paper_0269/figures/fig_6_p6/crops/bbox_27_ours.png", "paper_0269/figures/fig_6_p6/crops/bbox_16_legoformer.png" ]
Detail
[ "Object Form", "Detail" ]
our model produces reconstructions with the highest level of detail
[]
Figure 6. Examples of reconstructions produced by our method along with prior work on seven classes from the ShapeNet-R2N2 dataset [1]. The leftmost image in each row is the conditioning image. Given the simplicity and synthetic nature of ShapeNet, all models produce adequate results; however, our model produces recons...
[]
highest level of detail in point cloud reconstructions
our model produces reconstructions with the highest level of detail
image B model produces reconstructions with the highest level of detail
Object Form
Detail
0.82
paper_0269/fig_6_p6::row_3
paper_0541__fig_4_p6__b
paper_0541
fig_4_p6
b
b
Which method demonstrates better Ability to correctly reassemble puzzle pieces with missing patches and mainta...?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "Gallagher", "Yu et al.", "Ours", "Huroyan et al." ]
[ "paper_0541/figures/fig_4_p6/crops/bbox_04_gallagher.png", "paper_0541/figures/fig_4_p6/crops/bbox_06_yu_et_al.png", "paper_0541/figures/fig_4_p6/crops/bbox_01_ours.png", "paper_0541/figures/fig_4_p6/crops/bbox_02_huroyan_et_al.png" ]
Completeness
[ "Object Form", "Completeness" ]
Qualitative comparison of jigsaw puzzle solving with 30% missing pieces across different methods
[]
Figure 4. (a) Patch alignment in inference from t = T to t = 0. (b) Qualitative comparison with 30% missing pieces. cessfully placed if it is both correctly positioned and rotated. While our solution operates in the continuous domain, the evaluation is conducted in the discrete one of [14]. To do so, we first apply to e...
[]
Ability to correctly reassemble puzzle pieces with missing patches and mainta...
Qualitative comparison of jigsaw puzzle solving with 30% missing pieces across different methods
null
Object Form
Completeness
0.72
paper_0541/fig_4_p6::b
paper_1286__fig_10_p7__row_1
paper_1286
fig_10_p7
row_1
row_1
Which method demonstrates better Consistent reconstruction quality across different illumination conditions wi...?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "AvatarMe++", "Dib et al." ]
[ "paper_1286/figures/fig_10_p7/crops/bbox_00_ours.png", "paper_1286/figures/fig_10_p7/crops/bbox_08_avatarme_plus_plus.png", "paper_1286/figures/fig_10_p7/crops/bbox_04_dib_et_al.png" ]
Reconstruction Fidelity
[ "Reference Fidelity", "Reconstruction Fidelity" ]
Relightify produces superior reconstructions under different illumination conditions compared to Dib et al. and AvatarMe++, with better recovery of diffuse and specular albedo
[]
Figure 10. Reconstructions from images with different illumina- tion (Digital Emily Project [1]) by our method as well as [17, 41] and ground truth. We show the diffuse and specular albedo for all methods (where available), plus the recovered texture for our method. 4.5. Reflectance Prediction To further assess our met...
[]
Consistent reconstruction quality across different illumination conditions wi...
Relightify produces superior reconstructions under different illumination conditions compared to Dib et al. and AvatarMe++, with better recovery of diffuse and specular albedo
image A produces superior reconstructions under different illumination conditions compared to Dib et al. and AvatarMe++, with better recovery of diffuse and specular albedo
Reference Fidelity
Reconstruction Fidelity
0.88
paper_1286/fig_10_p7::row_1
paper_0450__fig_3_p5__row_2
paper_0450
fig_3_p5
row_2
row_2
Which method demonstrates better higher rendering quality with better detail preservation around occlusions an...?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours (141.7 FPS)", "KPlanes (1.4 FPS)", "ENeRF (11.3 FPS)" ]
[ "paper_0450/figures/fig_3_p5/crops/bbox_01_ours.png", "paper_0450/figures/fig_3_p5/crops/bbox_02_kplanes.png", "paper_0450/figures/fig_3_p5/crops/bbox_06_enerf.png" ]
Detail
[ "Object Form", "Detail" ]
Our method achieves much higher rendering quality and can be rendered 14× faster than ENeRF on the ENeRF-Outdoor dataset
[]
Figure 3. Qualitative comparison on the ENeRF-Outdoor [49] dataset that contains 960 × 540 images. Our method achieves much higher rendering quality and can be rendered 14× faster than ENeRF[49]. More dynamic results can be found in the supplementary video. only are there multiple moving humans and objects, but the bac...
[]
higher rendering quality with better detail preservation around occlusions an...
Our method achieves much higher rendering quality and can be rendered 14× faster than ENeRF on the ENeRF-Outdoor dataset
image A method achieves much higher rendering quality and can be rendered 14× faster than ENeRF on the ENeRF-Outdoor dataset
Object Form
Detail
0.7
paper_0450/fig_3_p5::row_2
paper_1109__fig_4_p4__the_great_wave
paper_1109
fig_4_p4
the_great_wave
the_great_wave
Which method demonstrates better Minimal interference with non-target artistic styles while effectively erasin...?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "Self Attentions", "ESD-u", "All Layers", "ESD-x" ]
[ "paper_1109/figures/fig_4_p4/crops/bbox_19_self_attentions.png", "paper_1109/figures/fig_4_p4/crops/bbox_23_esd_u.png", "paper_1109/figures/fig_4_p4/crops/bbox_00_all_layers.png", "paper_1109/figures/fig_4_p4/crops/bbox_03_esd_x.png" ]
Semantic Match
[ "Prompt Match", "Semantic Match" ]
ESD-x shows negligible interference with other styles and is well-suited for erasing art styles, while ESD-u has a global erasure effect and is better suited for removing nudity or objects
[]
Figure 4: Modifying the cross-attention weights, ESD-x, shows negligible interference with other styles (bottom 3 rows) and is thus well-suited for erasing art styles. In con- trast, altering the non-cross-attention weights, ESD-u, has a global erasure effect (all rows) on the visual concept and is better suited for re...
[]
Minimal interference with non-target artistic styles while effectively erasin...
ESD-x shows negligible interference with other styles and is well-suited for erasing art styles, while ESD-u has a global erasure effect and is better suited for removing nudity or objects
image D shows negligible interference with other styles and is well-suited for erasing art styles, while ESD-u has a global erasure effect and is better suited for removing nudity or objects
Prompt Match
Semantic Match
0.78
paper_1109/fig_4_p4::the_great_wave
paper_0452__fig_5_p6__row_4
paper_0452
fig_5_p6
row_4
row_4
Which method demonstrates better Hand mesh reconstruction accuracy and completeness compared to MobRecon, METR...?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "MeshGraphormer [11]", "Ours", "METRO [27]" ]
[ "paper_0452/figures/fig_5_p6/crops/bbox_13_meshgraphormer.png", "paper_0452/figures/fig_5_p6/crops/bbox_09_ours.png", "paper_0452/figures/fig_5_p6/crops/bbox_01_metro.png" ]
Correspondence
[ "Relation", "Correspondence" ]
Qualitative comparison between our method and other state-of-the-art approaches for hand mesh reconstruction
[]
Figure 5. Qualitative comparison between our method and other state-of-the-art approaches. sor is designed as a Multilayer Perceptron (MLP). We sub- sequently substitute these components with our proposed structures individually. The results of these experiments, detailed in Table 4, confirm that both modules, when inc...
[]
Hand mesh reconstruction accuracy and completeness compared to MobRecon, METR...
Qualitative comparison between our method and other state-of-the-art approaches for hand mesh reconstruction
Qualitative comparison between image B method and other state-of-the-art approaches for hand mesh reconstruction
Relation
Correspondence
0.72
paper_0452/fig_5_p6::row_4
paper_0085__fig_4_p5__scene_2
paper_0085
fig_4_p5
scene_2
scene_2
Which method demonstrates better Camera trajectory accuracy and drift reduction?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "ESLAM (ours)", "iMAP*", "NICE-SLAM" ]
[ "paper_0085/figures/fig_4_p5/crops/bbox_04_eslam.png", "paper_0085/figures/fig_4_p5/crops/bbox_03_imap_star.png", "paper_0085/figures/fig_4_p5/crops/bbox_07_nice_slam.png" ]
Landmark Fidelity
[ "Reference Fidelity", "Landmark Fidelity" ]
ESLAM predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on ScanNet dataset
[]
Figure 4. Qualitative comparison of our proposed ESLAM method’s localization accuracy with existing NeRF-based dense visual SLAM models, iMAP∗[59] and NICE-SLAM [87], on the ScanNet dataset [12]. The ground truth camera trajectory is shown in green, and the estimated trajectory is shown in red. Our method predicts more...
[]
Camera trajectory accuracy and drift reduction
ESLAM predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on ScanNet dataset
image A predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on ScanNet dataset
Reference Fidelity
Landmark Fidelity
0.82
paper_0085/fig_4_p5::scene_2
paper_0608__fig_9_p7__train
paper_0608
fig_9_p7
train
train
Which method demonstrates better modeling of tiny and sharp edges?
[ "(A)", "(B)" ]
(A)
true
[ "GES(ours)", "Gaussians" ]
[ "paper_0608/figures/fig_9_p7/crops/bbox_04_ges_ours.png", "paper_0608/figures/fig_9_p7/crops/bbox_00_gaussians.png" ]
Shape
[ "Object Form", "Shape" ]
GES can better model tiny and sharp edges compared to Gaussians when constrained to the same number of splatting components
[]
Figure 9. A Fair Visual Comparison. We show an example of Gaussians [18] and GES when constrained to the same number of splatting components for a fair visual comparison. It clearly shows that GES can better model tiny and sharp edges for that scene. 6. Results 6.1. Novel View Synthesis Results We evaluated GES against...
[]
modeling of tiny and sharp edges
GES can better model tiny and sharp edges compared to Gaussians when constrained to the same number of splatting components
image A can better model tiny and sharp edimage A compared to Gaussians when constrained to the same number of splatting components
Object Form
Shape
0.84
paper_0608/fig_9_p7::train
paper_1286__fig_5_p5__row_2
paper_1286
fig_5_p5
row_2
row_2
Which method demonstrates better Completeness and accuracy of reconstructed facial shape and reflectance prope...?
[ "(A)", "(B)", "(C)" ]
(C)
true
[ "AvatarMe++", "AlbedoMM", "Ours" ]
[ "paper_1286/figures/fig_5_p5/crops/bbox_13_avatarme_plus_plus.png", "paper_1286/figures/fig_5_p5/crops/bbox_11_albedomm.png", "paper_1286/figures/fig_5_p5/crops/bbox_06_ours.png" ]
Reconstruction Fidelity
[ "Reference Fidelity", "Reconstruction Fidelity" ]
Rendered reconstructions of shape and reflectance show superior quality compared to baseline methods
[]
Figure 5. Rendered reconstructions of shape and reflectance by AlbedoMM [60] (using the open-source code), AvatarMe++ [41] (provided by authors) and our method, in the same illumination. bitrary realistic environments, directly on the UV space: ρ(AD, AS, N, S) →T. For an evaluation of the model without re-rendered text...
[]
Completeness and accuracy of reconstructed facial shape and reflectance prope...
Rendered reconstructions of shape and reflectance show superior quality compared to baseline methods
null
Reference Fidelity
Reconstruction Fidelity
0.65
paper_1286/fig_5_p5::row_2
paper_0968__fig_3_p4__row_3
paper_0968
fig_3_p4
row_3
row_3
Which method demonstrates better Darker error maps indicating better depth estimation uniformly across image?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "Adabins", "Image", "Ours", "Ours Error" ]
[ "paper_0968/figures/fig_3_p4/crops/bbox_06_adabins.png", "paper_0968/figures/fig_3_p4/crops/bbox_19_image.png", "paper_0968/figures/fig_3_p4/crops/bbox_13_ours.png", "paper_0968/figures/fig_3_p4/crops/bbox_22_ours_error.png" ]
Reconstruction Fidelity
[ "Reference Fidelity", "Reconstruction Fidelity" ]
WorDepth improves uniformly across the image (darker in error map), implying better scale. WorDepth also predicts more accurate depth in regions corresponding to chairs, window, shower curtain, man, and desks, which are objects specified by text descriptions.
[]
Figure 3. Qualitative results on NYU Depth V2. We compare WorDepth with AdaBins [2]. Text descriptions are generated using ExpansionNet v2 [25]. Overall, WorDepth improves uniformly across the image (darker in error map), implying better scale. WorDepth also predicts more accurate depth in regions corresponding to “cha...
[]
Darker error maps indicating better depth estimation uniformly across image
WorDepth improves uniformly across the image (darker in error map), implying better scale. WorDepth also predicts more accurate depth in regions corresponding to chairs, window, shower curtain, man, and desks, which are objects specified by text descriptions.
image C improves uniformly across the image (darker in error map), implying better scale. image C also predicts more accurate depth in regions corresponding to chairs, window, shower curtain, man, and desks, which are objects specified by text descriptions.
Reference Fidelity
Reconstruction Fidelity
0.85
paper_0968/fig_3_p4::row_3
paper_0607__fig_3_p6__twindom
paper_0607
fig_3_p6
twindom
twindom
Which method demonstrates better detailed human appearances?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "ENeRF", "IBRNet", "Ours", "FloRen" ]
[ "paper_0607/figures/fig_3_p6/crops/bbox_23_enerf.png", "paper_0607/figures/fig_3_p6/crops/bbox_11_ibrnet.png", "paper_0607/figures/fig_3_p6/crops/bbox_35_ours.png", "paper_0607/figures/fig_3_p6/crops/bbox_25_floren.png" ]
Texture
[ "Object Form", "Texture" ]
GPS-Gaussian produces more detailed human appearances and can recover more reasonable geometry compared to baselines (ENeRF, IBRNet, FloRen, 3D-GS) on THuman2.0, Twindom, and real-world data
[]
Figure 3. Qualitative comparison on THuman2.0 [69], Twindom [55] and our collected real-world data. Our method produces more detailed human appearances and can recover more reasonable geometry. difficulty in handling thin structures such as hockey sticks and robes in Fig. 3. We further prepare the sensitivity analy- si...
[]
detailed human appearances
GPS-Gaussian produces more detailed human appearances and can recover more reasonable geometry compared to baselines (ENeRF, IBRNet, FloRen, 3D-GS) on THuman2.0, Twindom, and real-world data
image C produces more detailed human appearances and can recover more reasonable geometry compared to baselines (ENeRF, IBRNet, FloRen, 3D-GS) on THuman2.0, Twindom, and real-world data
Object Form
Texture
0.79
paper_0607/fig_3_p6::twindom
paper_1132__fig_8_p6__frame_15
paper_1132
fig_8_p6
frame_15
frame_15
Which method demonstrates better Ability to synthesize plausible novel views in occluded regions and handle am...?
[ "(A)", "(B)" ]
(A)
true
[ "Ours (One Step)", "Ours" ]
[ "paper_1132/figures/fig_8_p6/crops/bbox_09_ours_one_step.png", "paper_1132/figures/fig_8_p6/crops/bbox_10_ours.png" ]
Composition
[ "Scene Layout", "Composition" ]
Generative sampling with autoregressive generation enables plausible synthesis in ambiguity and explores areas that were completely occluded in the input, outperforming regression-based models and LOTR
[]
Figure 8. Regression-based models, such as the one-step variant of our approach, struggle to model ambiguity and therefore fail to create plausible renderings far from the input. Generative sampling enables plausible synthesis in ambiguity. When combined with autoregressive generation, we are able to explore areas that...
[]
Ability to synthesize plausible novel views in occluded regions and handle am...
Generative sampling with autoregressive generation enables plausible synthesis in ambiguity and explores areas that were completely occluded in the input, outperforming regression-based models and LOTR
null
Scene Layout
Composition
0.75
paper_1132/fig_8_p6::frame_15
paper_1304__fig_4_p5__full
paper_1304
fig_4_p5
full
full
Which method demonstrates better Reduced fog artifacts and improved scene clarity in rendered outputs?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "NSG", "ScatterNSG (descattered)", "ScatterNSG", "Original" ]
[ "paper_1304/figures/fig_4_p5/crops/bbox_00_nsg.png", "paper_1304/figures/fig_4_p5/crops/bbox_11_scatter_nsg_descattered.png", "paper_1304/figures/fig_4_p5/crops/bbox_08_scatter_nsg.png", "paper_1304/figures/fig_4_p5/crops/bbox_07_original.png" ]
Global Artifact
[ "Image Appearance", "Global Artifact" ]
ScatterNSG produces better visual quality compared to NSG baseline in fog removal and scene reconstruction
[]
Figure 4: Qualitative comparison between Neural Scene Graphs (NSG) [47] and its combination with our proposed network (ScatterNSG). 5.1. Ablation In order to assess the role and contribution of the differ- ent components of our framework, we conduct an ablation study whose results are presented in Tab. 3. We consider
[]
Reduced fog artifacts and improved scene clarity in rendered outputs
ScatterNSG produces better visual quality compared to NSG baseline in fog removal and scene reconstruction
image C produces better visual quality compared to NSG baseline in fog removal and scene reconstruction
Image Appearance
Global Artifact
0.79
paper_1304/fig_4_p5::full
paper_0318__fig_5_p6__sample_2
paper_0318
fig_5_p6
sample_2
sample_2
Which method demonstrates better Reduced noise and improved consistency across views compared to NVOS?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "Ours (Two-Stage)", "Ours", "NVOS", "Video Segmentation (2D)" ]
[ "paper_0318/figures/fig_5_p6/crops/bbox_03_ours_two_stage.png", "paper_0318/figures/fig_5_p6/crops/bbox_02_ours.png", "paper_0318/figures/fig_5_p6/crops/bbox_10_nvos.png", "paper_0318/figures/fig_5_p6/crops/bbox_09_video_segmentation_2d.png" ]
Boundary
[ "Object Form", "Boundary" ]
Our method is less noisy than NVOS, which also misses some pieces of the target, but captures more details than the video segmentation alone
[]
Figure 5. A qualitative comparison of our multiview segmentation model against Neural Volumetric Object Selection (NVOS) [44] (re- produced from original paper), video segmentation [4], and the human-annotated masks (GT). Our two-stage approach refers to running our multiview segmentation model twice, with the output o...
[]
Reduced noise and improved consistency across views compared to NVOS
Our method is less noisy than NVOS, which also misses some pieces of the target, but captures more details than the video segmentation alone
image B method is less noisy than NVOS, which also misses some pieces of the target, but captures more details than the video segmentation alone
Object Form
Boundary
0.75
paper_0318/fig_5_p6::sample_2
paper_1299__fig_7_p6__chair_sample
paper_1299
fig_7_p6
chair_sample
chair_sample
Which method demonstrates better Cleaner shapes with detailed structures and sharp edges on sparse real-world ...?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "PoinTr", "Ours", "Partial", "SeedFormer" ]
[ "paper_1299/figures/fig_7_p6/crops/bbox_01_pointr.png", "paper_1299/figures/fig_7_p6/crops/bbox_02_ours.png", "paper_1299/figures/fig_7_p6/crops/bbox_05_partial.png", "paper_1299/figures/fig_7_p6/crops/bbox_10_seedformer.png" ]
Boundary
[ "Object Form", "Boundary" ]
SVDFormer produces cleaner shapes with detailed structures and sharp edges compared to PoinTr and SeedFormer on real-world sparse point cloud scans
[]
Figure 7. Visual comparison on real-world scans. All results are produced by models pre-trained on PCN. Table 5. Effect of SVFNet. (ℓ1 CD ×103 and F1-Score@1%) Methods CD↓ DCD↓ F1↑
[]
Cleaner shapes with detailed structures and sharp edges on sparse real-world ...
SVDFormer produces cleaner shapes with detailed structures and sharp edges compared to PoinTr and SeedFormer on real-world sparse point cloud scans
image B produces cleaner shapes with detailed structures and sharp edges compared to PoinTr and SeedFormer on real-world sparse point cloud scans
Object Form
Boundary
0.81
paper_1299/fig_7_p6::chair_sample
paper_0815__fig_5_p5__frog_wizard
paper_0815
fig_5_p5
frog_wizard
frog_wizard
Which method demonstrates better Generation quality and semantic coherence of generated images across diverse ...?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "Stable Diffusion", "PRDP", "DDPO" ]
[ "paper_0815/figures/fig_5_p5/crops/bbox_16_stable_diffusion.png", "paper_0815/figures/fig_5_p5/crops/bbox_08_prdp.png", "paper_0815/figures/fig_5_p5/crops/bbox_13_ddpo.png" ]
Semantic Match
[ "Prompt Match", "Semantic Match" ]
PRDP significantly improves the generation quality over Stable Diffusion, whereas DDPO fails to generate reasonable results.
[]
Figure 5. Generation samples from large-scale training. DDPO and PRDP are finetuned from Stable Diffusion v1.4 on over 100K prompts from the training set of HPDv2, with HPSv2 (Left) and PickScore (Right) as the reward model. Samples within each column are generated from the prompt shown on top, using the same random se...
[]
Generation quality and semantic coherence of generated images across diverse ...
PRDP significantly improves the generation quality over Stable Diffusion, whereas DDPO fails to generate reasonable results.
image B significantly improves the generation quality over Stable Diffusion, whereas DDPO fails to generate reasonable results.
Prompt Match
Semantic Match
0.89
paper_0815/fig_5_p5::frog_wizard
paper_1123__fig_1_p0__time_step_2
paper_1123
fig_1_p0
time_step_2
time_step_2
Which method demonstrates better Accuracy of hand pose estimation under occlusion and blur conditions?
[ "(A)", "(B)" ]
(A)
true
[ "Ours", "TCMR" ]
[ "paper_1123/figures/fig_1_p0/crops/bbox_05_ours.png", "paper_1123/figures/fig_1_p0/crops/bbox_02_tcmr.png" ]
Pose
[ "Object Form", "Pose" ]
Given a video where in some frames the hand is heavily occluded or blurred, the existing state-of-the-art video-based method TCMR fails to predict accurate hand poses. Deformer is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible hand pose estimations.
[]
Figure 1: Given a video where in some frames (left) the hand is heavily occluded or blurred, the existing state-of- the-art video-based method TCMR [10] (middle) fails to predict accurate hand poses. Our method (right) is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible h...
[]
Accuracy of hand pose estimation under occlusion and blur conditions
Given a video where in some frames the hand is heavily occluded or blurred, the existing state-of-the-art video-based method TCMR fails to predict accurate hand poses. Deformer is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible hand pose estimations.
Given a video where in some frames the hand is heavily occluded or blurred, the existing state-of-the-art video-based method TCMR fails to predict accurate hand poses. image A is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible hand pose estimations.
Object Form
Pose
0.78
paper_1123/fig_1_p0::time_step_2
paper_0246__fig_4_p5__subject_2
paper_0246
fig_4_p5
subject_2
subject_2
Which method demonstrates better richer skin details and high-fidelity highlights?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "NeuFace", "VolSDF", "DIFFREC", "Ref-NeRF" ]
[ "paper_0246/figures/fig_4_p5/crops/bbox_02_neuface.png", "paper_0246/figures/fig_4_p5/crops/bbox_10_volsdf.png", "paper_0246/figures/fig_4_p5/crops/bbox_19_diffrec.png", "paper_0246/figures/fig_4_p5/crops/bbox_04_ref_nerf.png" ]
Texture
[ "Object Form", "Texture" ]
NeuFace captures faces with much richer skin details and high-fidelity highlights compared to Ref-NeRF, PhySG, VolSDF, and DIFFREC in novel view synthesis
[]
Figure 4. Comparison of novel view synthesis with Ref-NeRF [47], PhySG [56], VolSDF [54] and DIFFREC [31]. Zoom in for a better view. NeuFace captures faces with much richer skin details and high-fidelity highlights. w/ our calibration F. S.N. ↓ Subject 1 (Rich Reflection) Subject 2 (Moderate Reflection)
[]
richer skin details and high-fidelity highlights
NeuFace captures faces with much richer skin details and high-fidelity highlights compared to Ref-NeRF, PhySG, VolSDF, and DIFFREC in novel view synthesis
image A captures faces with much richer skin details and high-fidelity highlights compared to Ref-NeRF, PhySG, VolSDF, and DIFFREC in novel view synthesis
Object Form
Texture
0.79
paper_0246/fig_4_p5::subject_2
paper_0414__fig_4_p6__row_1
paper_0414
fig_4_p6
row_1
row_1
Which method demonstrates better Smooth boundaries and consistent room segmentation without aliasing artifacts...?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "(b) Label", "(f) DFPR", "(c)Ours", "(e) OCRNet" ]
[ "paper_0414/figures/fig_4_p6/crops/bbox_16_label.png", "paper_0414/figures/fig_4_p6/crops/bbox_00_dfpr.png", "paper_0414/figures/fig_4_p6/crops/bbox_17_ours.png", "paper_0414/figures/fig_4_p6/crops/bbox_23_ocrnet.png" ]
Global Artifact
[ "Image Appearance", "Global Artifact" ]
Our method generates smooth boundaries and consistent room segmentation, outperforming image-based methods that suffer from aliasing boundaries, inconsistent segmentation, and sawteeth patterns in predicted boundaries.
[]
Figure 4. The qualitative comparison with image-based floorplan segmentation methods on two datasets. The first two rows are examples from R2V and the last row is from Cubicasa. Vertex Embedding Architecture mIoU mAcc RI
[]
Smooth boundaries and consistent room segmentation without aliasing artifacts...
Our method generates smooth boundaries and consistent room segmentation, outperforming image-based methods that suffer from aliasing boundaries, inconsistent segmentation, and sawteeth patterns in predicted boundaries.
image C method generates smooth boundaries and consistent room segmentation, outperforming image-based methods that suffer from aliasing boundaries, inconsistent segmentation, and sawteeth patterns in predicted boundaries.
Image Appearance
Global Artifact
0.72
paper_0414/fig_4_p6::row_1
paper_1171__fig_5_p7__scene_1
paper_1171
fig_5_p7
scene_1
scene_1
Which method demonstrates better Fewer identity switches in multi-person tracking?
[ "(A)", "(B)" ]
(B)
true
[ "ByteTrack", "Egoformer (Ours)" ]
[ "paper_1171/figures/fig_5_p7/crops/bbox_00_bytetrack.png", "paper_1171/figures/fig_5_p7/crops/bbox_03_egoformer.png" ]
Correspondence
[ "Relation", "Correspondence" ]
EgoFormer performs identity association in 3D and suffers from fewer identity switches compared to ByteTrack
[]
Figure 5: Qualitative results of EgoFormer for tracking. In comparison to a 2D approach like ByteTrack [125], our proposed method performs identity association in 3D and suffers from fewer identity switches. Predicted person ids are shown in color. Method IDF1↑ HOTA↑ MOTA↑ IDs↓
[]
Fewer identity switches in multi-person tracking
EgoFormer performs identity association in 3D and suffers from fewer identity switches compared to ByteTrack
image B performs identity association in 3D and suffers from fewer identity switches compared to ByteTrack
Relation
Correspondence
0.82
paper_1171/fig_5_p7::scene_1
paper_1225__fig_5_p5__reception
paper_1225
fig_5_p5
reception
reception
Which method demonstrates better fewer artifacts?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "NICE-SLAM", "Ours", "ElasticFusion" ]
[ "paper_1225/figures/fig_5_p5/crops/bbox_10_nice_slam.png", "paper_1225/figures/fig_5_p5/crops/bbox_08_ours.png", "paper_1225/figures/fig_5_p5/crops/bbox_13_elasticfusion.png" ]
Attribute Match
[ "Object Form", "Attribute Match" ]
Our approach produces high-quality reconstructions with smooth surfaces and fewer artifacts compared to other methods
[]
Figure 5. Qualitative Reconstruction Results. Compared to other methods, our approach produces high-quality reconstructions with smooth surfaces and fewer artifacts. We also conduct a global BA every Ng frame by adding all the past mapping frames to the optimization window and jointly optimize them to globally optimize...
[]
fewer artifacts
Our approach produces high-quality reconstructions with smooth surfaces and fewer artifacts compared to other methods
image B approach produces high-quality reconstructions with smooth surfaces and fewer artifacts compared to other methods
Object Form
Attribute Match
0.88
paper_1225/fig_5_p5::reception
paper_0256__fig_3_p4__scan_37
paper_0256
fig_3_p4
scan_37
scan_37
Which method demonstrates better surface accuracy and fidelity?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "NeuralWarp", "NeuS" ]
[ "paper_0256/figures/fig_3_p4/crops/bbox_05_ours.png", "paper_0256/figures/fig_3_p4/crops/bbox_13_neural_warp.png", "paper_0256/figures/fig_3_p4/crops/bbox_04_neus.png" ]
Surface
[ "Object Form", "Surface" ]
Neuralangelo produces more accurate and higher-fidelity surfaces compared to NeuS and NeuralWarp on the DTU benchmark
[]
Figure 3. Qualitative comparison on the DTU benchmark [11]. Neuralangelo produces more accurate and higher-fidelity surfaces. finer hash grids throughout optimization when ϵ decreases to their spatial size. The relearning process can thus be avoided to better capture the details. In practice, we also apply weight decay...
[]
surface accuracy and fidelity
Neuralangelo produces more accurate and higher-fidelity surfaces compared to NeuS and NeuralWarp on the DTU benchmark
image A produces more accurate and higher-fidelity surfaces compared to NeuS and NeuralWarp on the DTU benchmark
Object Form
Surface
0.84
paper_0256/fig_3_p4::scan_37
paper_0710__fig_8_p5__row_2
paper_0710
fig_8_p5
row_2
row_2
Which method demonstrates better Consistency and quality of generated novel views compared to reference-guided...?
[ "(A)", "(B)" ]
(A)
true
[ "LeftRefill", "Zero123" ]
[ "paper_0710/figures/fig_8_p5/crops/bbox_15_leftrefill.png", "paper_0710/figures/fig_8_p5/crops/bbox_01_zero123.png" ]
View
[ "Scene Layout", "View" ]
LeftRefill enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
[]
Figure 8. NVS results on Objaverse [10] (row1, 2) and Google Scanned Objects [12] (row3, 4) from a single reference image. CLIP. Since image features from CLIP contain high-level se- mantics, they fail to deal with the fine-grained Ref-inpainting as shown in Figure 6(e)(f). Though TransFill [68] achieves proper results...
[]
Consistency and quality of generated novel views compared to reference-guided...
LeftRefill enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
image A enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
Scene Layout
View
0.75
paper_0710/fig_8_p5::row_2
paper_0186__fig_1_p0__rgb_input
paper_0186
fig_1_p0
rgb_input
rgb_input
Which method demonstrates better Accurate segmentation of object boundaries, particularly distinguishing cabin...?
[ "(A)", "(B)" ]
(A)
true
[ "Ours", "Baseline" ]
[ "paper_0186/figures/fig_1_p0/crops/bbox_03_ours.png", "paper_0186/figures/fig_1_p0/crops/bbox_05_baseline.png" ]
Segmentation Fidelity
[ "Reference Fidelity", "Segmentation Fidelity" ]
LRPNet correctly segments the cabinet from the wall with the proposed effective receptive field
[]
Figure 1. Qualitative results of Effective Receptive Field (ERF) [45]. Left: the input and ground truth. Middle: the ERF and results of baseline described in section 3.3. Right: the ERF and results of LRPNet (ours). The stained areas around the posi- tions of interest (red dots) represent the range of receptive fields,...
[]
Accurate segmentation of object boundaries, particularly distinguishing cabin...
LRPNet correctly segments the cabinet from the wall with the proposed effective receptive field
image A correctly segments the cabinet from the wall with the proposed effective receptive field
Reference Fidelity
Segmentation Fidelity
0.93
paper_0186/fig_1_p0::rgb_input
paper_0269__fig_6_p6__row_7
paper_0269
fig_6_p6
row_7
row_7
Which method demonstrates better highest level of detail in point cloud reconstructions?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "Pix2Vox++", "Ours", "LegoFormer" ]
[ "paper_0269/figures/fig_6_p6/crops/bbox_14_pix2vox.png", "paper_0269/figures/fig_6_p6/crops/bbox_23_ours.png", "paper_0269/figures/fig_6_p6/crops/bbox_08_legoformer.png" ]
Detail
[ "Object Form", "Detail" ]
Our model produces reconstructions with the highest level of detail compared to LegoFormer and Pix2Vox++ on ShapeNet-R2N2 dataset, particularly on challenging categories such as airplane and rifle.
[]
Figure 6. Examples of reconstructions produced by our method along with prior work on seven classes from the ShapeNet-R2N2 dataset [1]. The leftmost image in each row is the conditioning image. Given the simplicity and synthetic nature of ShapeNet, all models produce adequate results; however, our model produces recons...
[]
highest level of detail in point cloud reconstructions
Our model produces reconstructions with the highest level of detail compared to LegoFormer and Pix2Vox++ on ShapeNet-R2N2 dataset, particularly on challenging categories such as airplane and rifle.
image B model produces reconstructions with the highest level of detail compared to LegoFormer and Pix2Vox++ on ShapeNet-R2N2 dataset, particularly on challenging categories such as airplane and rifle.
Object Form
Detail
0.82
paper_0269/fig_6_p6::row_7
paper_0269__fig_6_p6__row_1
paper_0269
fig_6_p6
row_1
row_1
Which method demonstrates better highest level of detail in point cloud reconstructions?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "LegoFormer", "Pix2Vox++" ]
[ "paper_0269/figures/fig_6_p6/crops/bbox_12_ours.png", "paper_0269/figures/fig_6_p6/crops/bbox_10_legoformer.png", "paper_0269/figures/fig_6_p6/crops/bbox_33_pix2vox.png" ]
Detail
[ "Object Form", "Detail" ]
our model produces reconstructions with the highest level of detail
[]
Figure 6. Examples of reconstructions produced by our method along with prior work on seven classes from the ShapeNet-R2N2 dataset [1]. The leftmost image in each row is the conditioning image. Given the simplicity and synthetic nature of ShapeNet, all models produce adequate results; however, our model produces recons...
[]
highest level of detail in point cloud reconstructions
our model produces reconstructions with the highest level of detail
image A model produces reconstructions with the highest level of detail
Object Form
Detail
0.82
paper_0269/fig_6_p6::row_1
paper_0452__fig_5_p6__row_3
paper_0452
fig_5_p6
row_3
row_3
Which method demonstrates better Hand mesh reconstruction accuracy and completeness compared to MobRecon, METR...?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "METRO [27]", "MobRecon [4]", "Ours", "MeshGraphormer [11]" ]
[ "paper_0452/figures/fig_5_p6/crops/bbox_00_metro.png", "paper_0452/figures/fig_5_p6/crops/bbox_12_mobrecon.png", "paper_0452/figures/fig_5_p6/crops/bbox_04_ours.png", "paper_0452/figures/fig_5_p6/crops/bbox_19_meshgraphormer.png" ]
Correspondence
[ "Relation", "Correspondence" ]
Qualitative comparison between our method and other state-of-the-art approaches for hand mesh reconstruction
[]
Figure 5. Qualitative comparison between our method and other state-of-the-art approaches. sor is designed as a Multilayer Perceptron (MLP). We sub- sequently substitute these components with our proposed structures individually. The results of these experiments, detailed in Table 4, confirm that both modules, when inc...
[]
Hand mesh reconstruction accuracy and completeness compared to MobRecon, METR...
Qualitative comparison between our method and other state-of-the-art approaches for hand mesh reconstruction
Qualitative comparison between image C method and other state-of-the-art approaches for hand mesh reconstruction
Relation
Correspondence
0.72
paper_0452/fig_5_p6::row_3
paper_0276__fig_6_p5__frame_40
paper_0276
fig_6_p5
frame_40
frame_40
Which method demonstrates better view-consistent recoloring results across multiple frames and viewpoints?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "Du et al. (Edit 1)", "Ours", "Du et al. (Edit 2)" ]
[ "paper_0276/figures/fig_6_p5/crops/bbox_10_du_et_al_edit_1.png", "paper_0276/figures/fig_6_p5/crops/bbox_06_ours.png", "paper_0276/figures/fig_6_p5/crops/bbox_00_du_et_al_edit_2.png" ]
Correspondence
[ "Relation", "Correspondence" ]
PaletteNeRF recolors the 3D scene consistently while maintaining photorealism across views, unlike Du et al. which either only modifies specific colors or introduces view-inconsistency
[]
Figure 6. Qualitative comparison with video palette-based re- coloring method [11]. Unlike Du et al. [11] that either only mod- ifies the lego color (Edit 1) or introduces view-inconsistency when also modifying the gloves (Edit 2), our method recolors the 3D scene consistently while maintaining photorealisim across vie...
[]
view-consistent recoloring results across multiple frames and viewpoints
PaletteNeRF recolors the 3D scene consistently while maintaining photorealism across views, unlike Du et al. which either only modifies specific colors or introduces view-inconsistency
image B recolors the 3D scene consistently while maintaining photorealism across views, unlike Du et al. which either only modifies specific colors or introduces view-inconsistency
Relation
Correspondence
0.78
paper_0276/fig_6_p5::frame_40
paper_1337__fig_6_p6__sample_3
paper_1337
fig_6_p6
sample_3
sample_3
Which method demonstrates better Quality of back reconstruction and overall avatar generation fidelity?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "PIFu", "ARCH++", "ARCH", "DINAR (Ours)" ]
[ "paper_1337/figures/fig_6_p6/crops/bbox_14_pifu.png", "paper_1337/figures/fig_6_p6/crops/bbox_12_arch_plus_plus.png", "paper_1337/figures/fig_6_p6/crops/bbox_11_arch.png", "paper_1337/figures/fig_6_p6/crops/bbox_05_dinar.png" ]
Completeness
[ "Object Form", "Completeness" ]
Qualitative comparison on the SnapshotPeople benchmark showing front and back views to evaluate the quality of back reconstruction across avatar creation methods
[]
Figure 6: Qualitative comparison on the SnapshotPeople benchmark. We compare our method with state-of-the-art approaches for creating avatars: PIFu, PHORHUM, ARCH, ARCH++, StylePeople. For each method, we show a front view and a back view to evaluate the quality of back reconstruction. Results for PHORHUM, ARCH and ARC...
[]
Quality of back reconstruction and overall avatar generation fidelity
Qualitative comparison on the SnapshotPeople benchmark showing front and back views to evaluate the quality of back reconstruction across avatar creation methods
null
Object Form
Completeness
0.78
paper_1337/fig_6_p6::sample_3
paper_1286__fig_5_p5__row_3
paper_1286
fig_5_p5
row_3
row_3
Which method demonstrates better Completeness and accuracy of reconstructed facial shape and reflectance prope...?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "AlbedoMM", "Ours", "AvatarMe++" ]
[ "paper_1286/figures/fig_5_p5/crops/bbox_01_albedomm.png", "paper_1286/figures/fig_5_p5/crops/bbox_04_ours.png", "paper_1286/figures/fig_5_p5/crops/bbox_10_avatarme_plus_plus.png" ]
Reconstruction Fidelity
[ "Reference Fidelity", "Reconstruction Fidelity" ]
Rendered reconstructions of shape and reflectance show superior quality compared to baseline methods
[]
Figure 5. Rendered reconstructions of shape and reflectance by AlbedoMM [60] (using the open-source code), AvatarMe++ [41] (provided by authors) and our method, in the same illumination. bitrary realistic environments, directly on the UV space: ρ(AD, AS, N, S) →T. For an evaluation of the model without re-rendered text...
[]
Completeness and accuracy of reconstructed facial shape and reflectance prope...
Rendered reconstructions of shape and reflectance show superior quality compared to baseline methods
null
Reference Fidelity
Reconstruction Fidelity
0.65
paper_1286/fig_5_p5::row_3
paper_0102__fig_6_p5__barn
paper_0102
fig_6_p5
barn
barn
Which method demonstrates better sharper detailed textures?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "H. et al.", "CCN-RF", "NPRG++", "Ours" ]
[ "paper_0102/figures/fig_6_p5/crops/bbox_27_h_et_al.png", "paper_0102/figures/fig_6_p5/crops/bbox_01_ccn_rf.png", "paper_0102/figures/fig_6_p5/crops/bbox_14_nprg++.png", "paper_0102/figures/fig_6_p5/crops/bbox_00_ours.png" ]
Texture
[ "Object Form", "Texture" ]
Our method renders sharper detailed textures
[]
Figure 6. Qualitative evaluations on Tanks and Temples [16] dataset. Our method renders sharper detailed textures. Implementation Details. There are two ways of raster- ization: PyTorch3D [34] for headless server and OpenGL [40] for real-time rendering. Since our training is performed on a headless server, our main res...
[]
sharper detailed textures
Our method renders sharper detailed textures
image D method renders sharper detailed textures
Object Form
Texture
0.87
paper_0102/fig_6_p5::barn
paper_0011__fig_7_p6__row_2
paper_0011
fig_7_p6
row_2
row_2
Which method demonstrates better ALTO achieves detailed qualitative results with higher fidelity reconstructio...?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "(f) ALTO", "(d) ConvONet [52]", "(e) POCO [4]", "(c) SPSR [37]" ]
[ "paper_0011/figures/fig_7_p6/crops/bbox_04_alto.png", "paper_0011/figures/fig_7_p6/crops/bbox_06_convonet.png", "paper_0011/figures/fig_7_p6/crops/bbox_17_poco.png", "paper_0011/figures/fig_7_p6/crops/bbox_05_spsr.png" ]
Detail
[ "Object Form", "Detail" ]
ALTO can reconstruct the (top scene) double-deck table and (bottom scene) details in the chair with higher fidelity reconstruction due to combining ideas from point and grid latents
[]
Figure 7. Qualitative comparison on scene-level reconstruction Synthetic Room Dataset. Learning-based methods are trained and tested on 10K noisy points. ALTO can reconstruct the (top scene) double-deck table and (bottom scene) details in the chair. Method IoU ↑ Chamfer-L1 ↓ NC↑ F-score↑
[]
ALTO achieves detailed qualitative results with higher fidelity reconstructio...
ALTO can reconstruct the (top scene) double-deck table and (bottom scene) details in the chair with higher fidelity reconstruction due to combining ideas from point and grid latents
image A can reconstruct the (top scene) double-deck table and (bottom scene) details in the chair with higher fidelity reconstruction due to combining ideas from point and grid latents
Object Form
Detail
0.81
paper_0011/fig_7_p6::row_2
paper_1044__fig_4_p7__building_a
paper_1044
fig_4_p7
building_a
building_a
Which method demonstrates better Wireframe and mesh model quality for roof structure reconstruction?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "PolyFit", "2.5D Dual contouring", "Our supervised", "Building point clouds" ]
[ "paper_1044/figures/fig_4_p7/crops/bbox_12_polyfit.png", "paper_1044/figures/fig_4_p7/crops/bbox_06_2.5d_dual_contouring.png", "paper_1044/figures/fig_4_p7/crops/bbox_17_our_supervised.png", "paper_1044/figures/fig_4_p7/crops/bbox_07_building_point_clouds.png" ]
Completeness
[ "Object Form", "Completeness" ]
Figure 4 shows visualization results comparing three traditional methods and the paper's proposed method for roof reconstruction from point clouds across different building types
[]
Figure 4. The visualization results from three traditional and our methods 5. Multi-purpose of Building3D Dataset Building Semantic & Part Segmentation Our building point clouds, roof point clouds and facade point clouds have been assigned unique labels when building the roof recon- struction dataset. Therefore, Buildi...
[]
Wireframe and mesh model quality for roof structure reconstruction
Figure 4 shows visualization results comparing three traditional methods and the paper's proposed method for roof reconstruction from point clouds across different building types
null
Object Form
Completeness
0.78
paper_1044/fig_4_p7::building_a
paper_0312__fig_6_p6__scene_3
paper_0312
fig_6_p6
scene_3
scene_3
Which method demonstrates better absence of dust artifacts and improved geometric detail recovery?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "SCADE", "DDP", "Vanilla" ]
[ "paper_0312/figures/fig_6_p6/crops/bbox_00_scade.png", "paper_0312/figures/fig_6_p6/crops/bbox_13_ddp.png", "paper_0312/figures/fig_6_p6/crops/bbox_17_vanilla.png" ]
Detail
[ "Object Form", "Detail" ]
SCADE is able to avoid producing clouds of dust that are present in baseline results and recover fine object details such as blinds, chair backs, and piano stool legs
[]
Figure 6. Qualitative results on ScanNet. As shown, SCADE is able to avoid producing the clouds of dust that are present in the results of baselines (1st, 3rd, last columns). Moreover, SCADE is also able to snap and recover objects in the scene such as the details on the blinds (2nd column), the back of the chair (4th ...
[]
absence of dust artifacts and improved geometric detail recovery
SCADE is able to avoid producing clouds of dust that are present in baseline results and recover fine object details such as blinds, chair backs, and piano stool legs
image A is able to avoid producing clouds of dust that are present in baseline results and recover fine object details such as blinds, chair backs, and piano stool legs
Object Form
Detail
0.75
paper_0312/fig_6_p6::scene_3
paper_0710__fig_8_p5__row_1
paper_0710
fig_8_p5
row_1
row_1
Which method demonstrates better Consistency and quality of generated novel views compared to reference-guided...?
[ "(A)", "(B)" ]
(A)
true
[ "LeftRefill", "Zero123" ]
[ "paper_0710/figures/fig_8_p5/crops/bbox_05_leftrefill.png", "paper_0710/figures/fig_8_p5/crops/bbox_10_zero123.png" ]
View
[ "Scene Layout", "View" ]
LeftRefill enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
[]
Figure 8. NVS results on Objaverse [10] (row1, 2) and Google Scanned Objects [12] (row3, 4) from a single reference image. CLIP. Since image features from CLIP contain high-level se- mantics, they fail to deal with the fine-grained Ref-inpainting as shown in Figure 6(e)(f). Though TransFill [68] achieves proper results...
[]
Consistency and quality of generated novel views compared to reference-guided...
LeftRefill enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
image A enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
Scene Layout
View
0.75
paper_0710/fig_8_p5::row_1
paper_0575__fig_1_p0__apple
paper_0575
fig_1_p0
apple
apple
Which method demonstrates better precise manipulation and detail preservation?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "DragGAN", "DragDiffusion\nw/o LoRA finetuning", "Ours\nw/ mask", "DragDiffusion" ]
[ "paper_0575/figures/fig_1_p0/crops/bbox_09_draggan.png", "paper_0575/figures/fig_1_p0/crops/bbox_11_dragdiffusion_wo_lora.png", "paper_0575/figures/fig_1_p0/crops/bbox_10_ours_w_mask.png", "paper_0575/figures/fig_1_p0/crops/bbox_06_dragdiffusion.png" ]
Shape
[ "Object Form", "Shape" ]
EasyDrag ensures precise manipulation and detail preservation without fine-tuning or masks, outperforming DragGAN, DragDiffusion, and SDE-Drag
[]
Figure 1. DragGAN fails in both cases due to limited model capacity. Similarly, DragDiffusion relies on LoRA fine-tuning and hand-drawn masks to achieve high-quality results, and SDE-Drag also fails to address these two cases effectively. In contrast, our approach ensures precise manipulation and detail preservation wi...
[]
precise manipulation and detail preservation
EasyDrag ensures precise manipulation and detail preservation without fine-tuning or masks, outperforming DragGAN, DragDiffusion, and SDE-Drag
image C ensures precise manipulation and detail preservation without fine-tuning or masks, outperforming DragGAN, DragDiffusion, and SDE-Drag
Object Form
Shape
0.75
paper_0575/fig_1_p0::apple
paper_0058__fig_4_p5__row_2
paper_0058
fig_4_p5
row_2
row_2
Which method demonstrates better Segmentation accuracy and correctness of class predictions in nighttime urban...?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "(a) Source", "(c) Visual Ref.", "(e) Ours", "(d) HRDA" ]
[ "paper_0058/figures/fig_4_p5/crops/bbox_07_source.png", "paper_0058/figures/fig_4_p5/crops/bbox_11_visual_ref.png", "paper_0058/figures/fig_4_p5/crops/bbox_03_ours.png", "paper_0058/figures/fig_4_p5/crops/bbox_06_hrda.png" ]
Segmentation
[ "Object Form", "Segmentation" ]
IR2F-RMM produces better semantic segmentation results compared to HRDA under Cityscapes to Dark Zurich domain adaptation
[]
Figure 4. Qualitative Comparisons for UDA Semantic Segmen- tation, under Cityscapes ! Dark Zurich. (a) shows the example of Cityscapes images. (b) includes the Dark Zurich images. (c) covers the day images corresponding to the night images in (b) for better visual references. Note that, the day images in (c) are only u...
[]
Segmentation accuracy and correctness of class predictions in nighttime urban...
IR2F-RMM produces better semantic segmentation results compared to HRDA under Cityscapes to Dark Zurich domain adaptation
null
Object Form
Segmentation
0.81
paper_0058/fig_4_p5::row_2
paper_0337__fig_6_p6__full_body
paper_0337
fig_6_p6
full_body
full_body
Which method demonstrates better hand articulation and facial expression accuracy?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "SCANimate", "Ours", "SNARF" ]
[ "paper_0337/figures/fig_6_p6/crops/bbox_12_scanimate.png", "paper_0337/figures/fig_6_p6/crops/bbox_13_ours.png", "paper_0337/figures/fig_6_p6/crops/bbox_05_snarf.png" ]
Anatomy
[ "Object Form", "Anatomy" ]
Our method recovers hand articulation and facial expression most accurately
[]
Figure 6. Qualitative results on GRAB dataset. Our method recovers hand articulation and facial expression most accurately. 4.4. Results on GRAB Dataset Tab. 2 summarizes results on the GRAB dataset. Over- all, our method beats all baselines, especially for the hands, where the margin is large. Fig. 6 visually shows th...
[]
hand articulation and facial expression accuracy
Our method recovers hand articulation and facial expression most accurately
image B method recovers hand articulation and facial expression most accurately
Object Form
Anatomy
0.82
paper_0337/fig_6_p6::full_body
paper_0444__fig_6_p6__main_scene
paper_0444
fig_6_p6
main_scene
main_scene
Which method demonstrates better Complete static scene reconstruction with correct preservation of static obje...?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "NKSR [18]", "SHINE-mapping [73]", "VDB-Fusion [59]", "Ours" ]
[ "paper_0444/figures/fig_6_p6/crops/bbox_00_nksr.png", "paper_0444/figures/fig_6_p6/crops/bbox_01_shine_mapping.png", "paper_0444/figures/fig_6_p6/crops/bbox_04_vdb_fusion.png", "paper_0444/figures/fig_6_p6/crops/bbox_03_ours.png" ]
Occlusion Layout
[ "Scene Layout", "Occlusion Layout" ]
Our method can reconstruct the static scene completely and eliminate the moving pedestrians. Although VDB-Fusion manages to eliminate the pedestrians, it incorrectly removes the tree highlighted in the orange box.
[]
Figure 6. A comparison of the static mapping results of different methods on the Newer College dataset. Several pedestrians are moving through the scene during the data collection. Our method can reconstruct the static scene completely and eliminate the moving pedestrians. Although VDB-Fusion manages to eliminate the p...
[]
Complete static scene reconstruction with correct preservation of static obje...
Our method can reconstruct the static scene completely and eliminate the moving pedestrians. Although VDB-Fusion manages to eliminate the pedestrians, it incorrectly removes the tree highlighted in the orange box.
image D method can reconstruct the static scene completely and eliminate the moving pedestrians. Although VDB-Fusion manages to eliminate the pedestrians, it incorrectly removes the tree highlighted in the orange box.
Scene Layout
Occlusion Layout
0.72
paper_0444/fig_6_p6::main_scene
paper_0981__fig_3_p5__row_1
paper_0981
fig_3_p5
row_1
row_1
Which method demonstrates better reconstruction of fine details such as beards and teeth?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "NeRFace", "Ours", "HAvatar", "NeRFBlendShape" ]
[ "paper_0981/figures/fig_3_p5/crops/bbox_17_nerf_face.png", "paper_0981/figures/fig_3_p5/crops/bbox_01_ours.png", "paper_0981/figures/fig_3_p5/crops/bbox_22_havatar.png", "paper_0981/figures/fig_3_p5/crops/bbox_12_nerf_blend_shape.png" ]
Detail
[ "Object Form", "Detail" ]
Our method can reconstruct details like beards, teeth, etc. with high quality.
[]
Figure 3. Qualitative comparisons of different methods on self reenactment task. From left to right: NeRFBlendShape [17], NeRFace [16], HAvatar [71] and Ours. Our method can reconstruct details like beards, teeth, etc. with high quality. cameras are distributed about 120 degrees in front, and si- multaneously capture 2...
[]
reconstruction of fine details such as beards and teeth
Our method can reconstruct details like beards, teeth, etc. with high quality.
image B method can reconstruct details like beards, teeth, etc. with high quality.
Object Form
Detail
0.85
paper_0981/fig_3_p5::row_1
paper_1157__fig_1_p0__row_1
paper_1157
fig_1_p0
row_1
row_1
Which method demonstrates better preservation of geometric structure and reduction of free-space voxels?
[ "(A)", "(B)" ]
(B)
true
[ "Previous work", "ImGeoNet" ]
[ "paper_1157/figures/fig_1_p0/crops/bbox_02_previous_work.png", "paper_1157/figures/fig_1_p0/crops/bbox_05_imgeonet.png" ]
Geometry Fidelity
[ "Reference Fidelity", "Geometry Fidelity" ]
ImGeoNet successfully preserves the geometric structure while effectively reducing the number of voxels in free space, in contrast to prior works that disregard the underlying geometry
[]
Figure 1. Geometry-aware voxel representation. (Top part) In contrast to prior works [56] (top left) that disregard the underly- ing geometry, our proposed ImGeoNet (top center) successfully preserves the geometric structure with respect to the ground truth (top right) while effectively reducing the number of voxels in...
[]
preservation of geometric structure and reduction of free-space voxels
ImGeoNet successfully preserves the geometric structure while effectively reducing the number of voxels in free space, in contrast to prior works that disregard the underlying geometry
image B successfully preserves the geometric structure while effectively reducing the number of voxels in free space, in contrast to prior works that disregard the underlying geometry
Reference Fidelity
Geometry Fidelity
0.85
paper_1157/fig_1_p0::row_1
paper_0275__fig_5_p6__row_5
paper_0275
fig_5_p6
row_5
row_5
Which method demonstrates better more realistic results?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "GVS* [16]", "Ours", "3DPhoto [56]", "InfiniteNatureZero [32]" ]
[ "paper_0275/figures/fig_5_p6/crops/bbox_15_gvs.png", "paper_0275/figures/fig_5_p6/crops/bbox_29_ours.png", "paper_0275/figures/fig_5_p6/crops/bbox_01_3dphoto.png", "paper_0275/figures/fig_5_p6/crops/bbox_03_infinitenaturezero.png" ]
Realism
[ "Image Appearance", "Realism" ]
Our method produces more realistic and view-consistent results compared to all baselines (3DPhoto, AdaMPI, GVS*, InfiniteNatureZero) on the LHQ dataset
[]
Figure 5. Qualitative comparisons on the LHQ dataset. We produce more realistic results compared to all baselines, as demonstrated in the supplementary video.
[]
more realistic results
Our method produces more realistic and view-consistent results compared to all baselines (3DPhoto, AdaMPI, GVS*, InfiniteNatureZero) on the LHQ dataset
image B method produces more realistic and view-consistent results compared to all baselines (3DPhoto, AdaMPI, GVS*, InfiniteNatureZero) on the LHQ dataset
Image Appearance
Realism
0.79
paper_0275/fig_5_p6::row_5
paper_1237__fig_6_p6__rendering
paper_1237
fig_6_p6
rendering
rendering
Which method demonstrates better Smoother normal estimation, more accurate environment map, better relighting ...?
[ "(A)", "(B)", "(C)" ]
(C)
true
[ "NeRFactor", "InvRender", "Ours" ]
[ "paper_1237/figures/fig_6_p6/crops/bbox_02_nerfactor.png", "paper_1237/figures/fig_6_p6/crops/bbox_15_invrender.png", "paper_1237/figures/fig_6_p6/crops/bbox_08_ours.png" ]
Surface
[ "Object Form", "Surface" ]
NeMF produces better rendering results with smoother normal estimation and more accurate environment map recovery compared to NeRFactor and InvRender, with improved relighting performance
[]
Figure 6: Comparisons with other approaches. We show the normal, albedo, roughness, and environment map estimated by NeRFactor, InvRender and our method on three scenes. We also compared the rendering results of the new viewpoint under three different lighting conditions (the original light and two novel lights). Note ...
[]
Smoother normal estimation, more accurate environment map, better relighting ...
NeMF produces better rendering results with smoother normal estimation and more accurate environment map recovery compared to NeRFactor and InvRender, with improved relighting performance
image C produces better rendering results with smoother normal estimation and more accurate environment map recovery compared to NeRFactor and InvRender, with improved relighting performance
Object Form
Surface
0.78
paper_1237/fig_6_p6::rendering
paper_0326__fig_9_p7__row_2
paper_0326
fig_9_p7
row_2
row_2
Which method demonstrates better Plane segmentation accuracy in edge regions?
[ "(A)", "(B)" ]
(B)
true
[ "Plane R-CNN", "Ours" ]
[ "paper_0326/figures/fig_9_p7/crops/bbox_09_plane_r_cnn.png", "paper_0326/figures/fig_9_p7/crops/bbox_07_ours.png" ]
Segmentation
[ "Object Form", "Segmentation" ]
Our method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN
[]
Figure 9. Plane Segmentation Comparisons. Our method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN [18]. Plane Reconstruction. To demonstrate the capability of our method on the unsupervised plane reconstruction task, we compare its reconstruction performance with the sta...
[]
Plane segmentation accuracy in edge regions
Our method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN
image B method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN
Object Form
Segmentation
0.75
paper_0326/fig_9_p7::row_2
paper_1123__fig_1_p0__time_step_4
paper_1123
fig_1_p0
time_step_4
time_step_4
Which method demonstrates better Accuracy of hand pose estimation under occlusion and blur conditions?
[ "(A)", "(B)" ]
(B)
true
[ "TCMR", "Ours" ]
[ "paper_1123/figures/fig_1_p0/crops/bbox_03_tcmr.png", "paper_1123/figures/fig_1_p0/crops/bbox_00_ours.png" ]
Pose
[ "Object Form", "Pose" ]
Given a video where in some frames the hand is heavily occluded or blurred, the existing state-of-the-art video-based method TCMR fails to predict accurate hand poses. Deformer is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible hand pose estimations.
[]
Figure 1: Given a video where in some frames (left) the hand is heavily occluded or blurred, the existing state-of- the-art video-based method TCMR [10] (middle) fails to predict accurate hand poses. Our method (right) is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible h...
[]
Accuracy of hand pose estimation under occlusion and blur conditions
Given a video where in some frames the hand is heavily occluded or blurred, the existing state-of-the-art video-based method TCMR fails to predict accurate hand poses. Deformer is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible hand pose estimations.
Given a video where in some frames the hand is heavily occluded or blurred, the existing state-of-the-art video-based method TCMR fails to predict accurate hand poses. image B is able to capture the hand dynamics and leverage neighborhood frames to robustly produce plausible hand pose estimations.
Object Form
Pose
0.78
paper_1123/fig_1_p0::time_step_4
paper_0143__fig_4_p4__row_3
paper_0143
fig_4_p4
row_3
row_3
Which method demonstrates better more credible identities close to the ground-truth?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "Low-Resolution", "Ground-Truth", "IDM (ours)", "SR3" ]
[ "paper_0143/figures/fig_4_p4/crops/bbox_08_low_resolution.png", "paper_0143/figures/fig_4_p4/crops/bbox_03_ground_truth.png", "paper_0143/figures/fig_4_p4/crops/bbox_16_idm_ours.png", "paper_0143/figures/fig_4_p4/crops/bbox_10_sr3.png" ]
Identity Preservation
[ "Reference Fidelity", "Identity Preservation" ]
IDM maintains higher fidelity and more credible identities close to the ground-truth, generating more realistic facial components (e.g., eyes, teeth, and hair) compared to SR3 on 8× SR on CelebA-HQ
[]
Figure 4. Qualitative comparison on 8× SR on CelebA-HQ [15]. The results of IDM maintain higher fidelity and more credible identities close to the ground-truth, generating more realistic facial components (e.g., eyes, teeth, and hair). 4.1. Implementation Details Datasets. We conduct our experiments on face datasets, n...
[]
more credible identities close to the ground-truth
IDM maintains higher fidelity and more credible identities close to the ground-truth, generating more realistic facial components (e.g., eyes, teeth, and hair) compared to SR3 on 8× SR on CelebA-HQ
image C maintains higher fidelity and more credible identities close to the ground-truth, generating more realistic facial components (e.g., eyes, teeth, and hair) compared to SR3 on 8× SR on CelebA-HQ
Reference Fidelity
Identity Preservation
0.9
paper_0143/fig_4_p4::row_3
paper_1263__fig_7_p7__row_2
paper_1263
fig_7_p7
row_2
row_2
Which method demonstrates better Ability to reconstruct 3D models from incomplete line drawings with only visi...?
[ "(A)", "(B)", "(C)" ]
(C)
true
[ "[25, 28]", "Front", "Ours" ]
[ "paper_1263/figures/fig_7_p7/crops/bbox_32_25_28.png", "paper_1263/figures/fig_7_p7/crops/bbox_02_front.png", "paper_1263/figures/fig_7_p7/crops/bbox_43_ours.png" ]
Completeness
[ "Object Form", "Completeness" ]
PlankAssembly correctly reconstructs objects from orthographic views with varying noise levels and incomplete inputs, while traditional methods fail or degrade significantly
[]
Figure 7. Qualitative results. Rows 1-2: clean inputs. Rows 3-5: Noisy inputs with noise level 10%, 20%, and 30%, respectively. Row 6: Inputs with visible parts only. We use red boxes to indicate incorrect reconstructions. 0 10 20 30 Noise Level (%)
[]
Ability to reconstruct 3D models from incomplete line drawings with only visi...
PlankAssembly correctly reconstructs objects from orthographic views with varying noise levels and incomplete inputs, while traditional methods fail or degrade significantly
image C correctly reconstructs objects from orthographic views with varying noise levels and incomplete inputs, while traditional methods fail or degrade significantly
Object Form
Completeness
0.78
paper_1263/fig_7_p7::row_2
paper_0710__fig_8_p5__row_4
paper_0710
fig_8_p5
row_4
row_4
Which method demonstrates better Consistency and quality of generated novel views compared to reference-guided...?
[ "(A)", "(B)" ]
(B)
true
[ "Zero123", "LeftRefill" ]
[ "paper_0710/figures/fig_8_p5/crops/bbox_14_zero123.png", "paper_0710/figures/fig_8_p5/crops/bbox_02_leftrefill.png" ]
View
[ "Scene Layout", "View" ]
LeftRefill enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
[]
Figure 8. NVS results on Objaverse [10] (row1, 2) and Google Scanned Objects [12] (row3, 4) from a single reference image. CLIP. Since image features from CLIP contain high-level se- mantics, they fail to deal with the fine-grained Ref-inpainting as shown in Figure 6(e)(f). Though TransFill [68] achieves proper results...
[]
Consistency and quality of generated novel views compared to reference-guided...
LeftRefill enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
image B enjoys substantial advantages in both qualitative and quantitative comparisons with negligible trainable weights in novel view synthesis from a single reference image
Scene Layout
View
0.75
paper_0710/fig_8_p5::row_4
paper_1075__fig_3_p4__row_1
paper_1075
fig_3_p4
row_1
row_1
Which method demonstrates better Improved depth map predictions and view synthesis quality from unseen viewpoints?
[ "(A)", "(B)", "(C)" ]
(C)
true
[ "Baseline (radiance field)", "DeLiRa (radiance field)", "DeLiRa (depth and light fields)" ]
[ "paper_1075/figures/fig_3_p4/crops/bbox_06_baseline_radiance.png", "paper_1075/figures/fig_3_p4/crops/bbox_12_delira_radiance.png", "paper_1075/figures/fig_3_p4/crops/bbox_02_delira_depth_light.png" ]
Detail
[ "Object Form", "Detail" ]
DeLiRa (depth and light fields) produces better depth and view synthesis results compared to baseline and DeLiRa (radiance field) variant
[]
Figure 3: Qualitative depth and view synthesis results from unseen viewpoints, using different DeLiRa decoders. As a baseline, we show predictions obtained from a model trained without our contributions, leading to a degenerate learned geometry due to shape-radiance ambiguity (i.e., accurate view synthesis with poor de...
[]
Improved depth map predictions and view synthesis quality from unseen viewpoints
DeLiRa (depth and light fields) produces better depth and view synthesis results compared to baseline and DeLiRa (radiance field) variant
image C produces better depth and view synthesis results compared to baseline and DeLiRa (radiance field) variant
Object Form
Detail
0.82
paper_1075/fig_3_p4::row_1
paper_0312__fig_7_p7__scene_2
paper_0312
fig_7_p7
scene_2
scene_2
Which method demonstrates better Recovery of objects behind glass and crisp shape preservation of thin structures?
[ "(A)", "(B)", "(C)" ]
(C)
true
[ "Vanilla", "DDP", "SCADE" ]
[ "paper_0312/figures/fig_7_p7/crops/bbox_05_vanilla.png", "paper_0312/figures/fig_7_p7/crops/bbox_22_ddp.png", "paper_0312/figures/fig_7_p7/crops/bbox_09_scade.png" ]
Detail
[ "Object Form", "Detail" ]
SCADE is able to recover the room behind the glass better than the baselines, whose results are more blurry. SCADE is able to recover crisp shapes of objects such as the thin table leg and the white chair, and is able to clear up dust near the microwave and printer compared to the baselines.
[]
Figure 7. Qualitative results on In-the-Wild data. As shown, we are able to recover the room behind the glass better the baselines, whose results are more blurry (1st, 3rd, 4th column). Also, SCADE is able to recover crisp shapes of objects such as the thin table leg and the white chair (2nd, last column), and we are a...
[]
Recovery of objects behind glass and crisp shape preservation of thin structures
SCADE is able to recover the room behind the glass better than the baselines, whose results are more blurry. SCADE is able to recover crisp shapes of objects such as the thin table leg and the white chair, and is able to clear up dust near the microwave and printer compared to the baselines.
image C is able to recover the room behind the glass better than the baselines, whose results are more blurry. image C is able to recover crisp shapes of objects such as the thin table leg and the white chair, and is able to clear up dust near the microwave and printer compared to the baselines.
Object Form
Detail
0.72
paper_0312/fig_7_p7::scene_2
paper_1141__fig_6_p6__row_3
paper_1141
fig_6_p6
row_3
row_3
Which method demonstrates better hair direction preservation?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "StyY.Hair", "Identity Hair", "HairNeRF", "Barbershop" ]
[ "paper_1141/figures/fig_6_p6/crops/bbox_17_styy_hair.png", "paper_1141/figures/fig_6_p6/crops/bbox_01_identity_hair.png", "paper_1141/figures/fig_6_p6/crops/bbox_03_hairnerf.png", "paper_1141/figures/fig_6_p6/crops/bbox_02_barbershop.png" ]
Shape
[ "Object Form", "Shape" ]
our method preserves the direction of the hair better
[]
Figure 6: Comparison of results for FFHQ datasets. Both StyleYourHair and our method align the hair even if the poses between the input images are different, but our method preserves the direction of the hair better (first and the last rows). ="GqiHwyxBQ7t2exuiSmBi8xrTEs=">AC1H icjVHLSsNAFD2Nr1ofjbp0EyCq5KIqMuiG91Vs A9...
[]
hair direction preservation
our method preserves the direction of the hair better
image C method preserves the direction of the hair better
Object Form
Shape
0.82
paper_1141/fig_6_p6::row_3
paper_1160__fig_6_p7__row_3
paper_1160
fig_6_p7
row_3
row_3
Which method demonstrates better Instance mask quality and multi-object segmentation accuracy?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "RGB", "Ours", "RFP", "uORF" ]
[ "paper_1160/figures/fig_6_p7/crops/bbox_12_rgb.png", "paper_1160/figures/fig_6_p7/crops/bbox_06_ours.png", "paper_1160/figures/fig_6_p7/crops/bbox_02_rfp.png", "paper_1160/figures/fig_6_p7/crops/bbox_08_uorf.png" ]
Mask Fidelity
[ "Reference Fidelity", "Mask Fidelity" ]
Instance Neural Radiance Field achieves superior instance segmentation quality compared to unsupervised methods (uORF, NeRF-SOS, RFP) with 95.48% accuracy and 89.34 mIoU on foreground/background segmentation
[]
Figure 6: Qualitative Results of Comparison with Unsuper- vised Methods. For unsupervised methods mentioned in section 2.3, none of them produces object classes. uORF [54] and ObSuRF [39] are completely unsupervised and work on object-centric scenes, which are far different from our test-
[]
Instance mask quality and multi-object segmentation accuracy
Instance Neural Radiance Field achieves superior instance segmentation quality compared to unsupervised methods (uORF, NeRF-SOS, RFP) with 95.48% accuracy and 89.34 mIoU on foreground/background segmentation
image B achieves superior instance segmentation quality compared to unsupervised methods (uORF, NeRF-SOS, RFP) with 95.48% accuracy and 89.34 mIoU on foreground/background segmentation
Reference Fidelity
Mask Fidelity
0.79
paper_1160/fig_6_p7::row_3
paper_0949__fig_5_p5__row_3
paper_0949
fig_5_p5
row_3
row_3
Which method demonstrates better Layout quality, visual coherence, and composition appeal?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "VLC (dual domain)", "VLC (vector only)", "Original Design" ]
[ "paper_0949/figures/fig_5_p5/crops/bbox_13_vlc_dual_domain.png", "paper_0949/figures/fig_5_p5/crops/bbox_04_vlc_vector_only.png", "paper_0949/figures/fig_5_p5/crops/bbox_03_original_design.png" ]
Completeness
[ "Object Form", "Completeness" ]
VLC (dual domain) achieves better results than VLC (vector only) and produces high-quality layouts that are more visually coherent and appealing
[]
Figure 5. Qualitative comparison of our method with baselines on Poster dataset. tential for generating high-quality layouts under varied and challenging conditions. 4.1.2 Qualitative Evaluation The effectiveness of our dual-domain method is further il-
[]
Layout quality, visual coherence, and composition appeal
VLC (dual domain) achieves better results than VLC (vector only) and produces high-quality layouts that are more visually coherent and appealing
image A achieves better results than VLC (vector only) and produces high-quality layouts that are more visually coherent and appealing
Object Form
Completeness
0.73
paper_0949/fig_5_p5::row_3
paper_0023__fig_3_p5__sample_2
paper_0023
fig_3_p5
sample_2
sample_2
Which method demonstrates better photorealistic image generation quality and visual fidelity?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "LDM", "CDE", "DRIT++", "BBDM(ours)" ]
[ "paper_0023/figures/fig_3_p5/crops/bbox_14_ldm.png", "paper_0023/figures/fig_3_p5/crops/bbox_02_cde.png", "paper_0023/figures/fig_3_p5/crops/bbox_00_drit_plus_plus.png", "paper_0023/figures/fig_3_p5/crops/bbox_03_bbdm_ours.png" ]
Realism
[ "Image Appearance", "Realism" ]
Qualitative comparison on CelebAMask-HQ dataset for semantic synthesis task
[]
Figure 3. Qualitative comparison on CelebAMask-HQ dataset. model and the proposed Brownian Bridge diffusion model. For fair comparison, we adopt the same pretrained VQGAN model as used in Latent Diffusion Model [28]. The number of time steps of Brownian Bridge is set to be 1000 during the training stage, and we use 200...
[]
photorealistic image generation quality and visual fidelity
Qualitative comparison on CelebAMask-HQ dataset for semantic synthesis task
null
Image Appearance
Realism
0.85
paper_0023/fig_3_p5::sample_2
paper_0931__fig_4_p6__row_1
paper_0931
fig_4_p6
row_1
row_1
Which method demonstrates better More detailed and clear boundaries in tampered region localization?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "Ours", "TruFor", "Forged", "CAT-Net V2" ]
[ "paper_0931/figures/fig_4_p6/crops/bbox_03_ours.png", "paper_0931/figures/fig_4_p6/crops/bbox_00_trufor.png", "paper_0931/figures/fig_4_p6/crops/bbox_36_forged.png", "paper_0931/figures/fig_4_p6/crops/bbox_29_cat_net_v2.png" ]
Boundary
[ "Object Form", "Boundary" ]
Our method can accurately locate tampered regions, predicting more detailed and clear boundaries
[]
Figure 4. Qualitative comparison results. The first to fourth rows are respectively sourced from CASIA v1 [14], Columbia [26], Coverage [61], and IMD20 [46]. The last row is from the BDNIE dataset. Variant Models CASIA v1 NIST 16 AUC F1
[]
More detailed and clear boundaries in tampered region localization
Our method can accurately locate tampered regions, predicting more detailed and clear boundaries
image A method can accurately locate tampered regions, predicting more detailed and clear boundaries
Object Form
Boundary
0.85
paper_0931/fig_4_p6::row_1
paper_0269__fig_6_p6__row_9
paper_0269
fig_6_p6
row_9
row_9
Which method demonstrates better highest level of detail in point cloud reconstructions?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "Pix2Vox++", "LegoFormer" ]
[ "paper_0269/figures/fig_6_p6/crops/bbox_04_ours.png", "paper_0269/figures/fig_6_p6/crops/bbox_31_pix2vox.png", "paper_0269/figures/fig_6_p6/crops/bbox_05_legoformer.png" ]
Detail
[ "Object Form", "Detail" ]
our model produces reconstructions with the highest level of detail
[]
Figure 6. Examples of reconstructions produced by our method along with prior work on seven classes from the ShapeNet-R2N2 dataset [1]. The leftmost image in each row is the conditioning image. Given the simplicity and synthetic nature of ShapeNet, all models produce adequate results; however, our model produces recons...
[]
highest level of detail in point cloud reconstructions
our model produces reconstructions with the highest level of detail
image A model produces reconstructions with the highest level of detail
Object Form
Detail
0.82
paper_0269/fig_6_p6::row_9
paper_0085__fig_4_p5__scene_1
paper_0085
fig_4_p5
scene_1
scene_1
Which method demonstrates better Camera trajectory accuracy and drift reduction?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "ESLAM (ours)", "NICE-SLAM", "iMAP*" ]
[ "paper_0085/figures/fig_4_p5/crops/bbox_06_eslam.png", "paper_0085/figures/fig_4_p5/crops/bbox_08_nice_slam.png", "paper_0085/figures/fig_4_p5/crops/bbox_05_imap_star.png" ]
Landmark Fidelity
[ "Reference Fidelity", "Landmark Fidelity" ]
ESLAM predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on the ScanNet dataset
[]
Figure 4. Qualitative comparison of our proposed ESLAM method’s localization accuracy with existing NeRF-based dense visual SLAM models, iMAP∗[59] and NICE-SLAM [87], on the ScanNet dataset [12]. The ground truth camera trajectory is shown in green, and the estimated trajectory is shown in red. Our method predicts more...
[]
Camera trajectory accuracy and drift reduction
ESLAM predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on the ScanNet dataset
image A predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on the ScanNet dataset
Reference Fidelity
Landmark Fidelity
0.82
paper_0085/fig_4_p5::scene_1
paper_0276__fig_7_p6__lego
paper_0276
fig_7_p6
lego
lego
Which method demonstrates better The recolored bonsai maintains consistent appearance across different viewing...?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "Tan et al.", "Ours", "Akimoto et al." ]
[ "paper_0276/figures/fig_7_p6/crops/bbox_23_tan_et_al.png", "paper_0276/figures/fig_7_p6/crops/bbox_10_ours.png", "paper_0276/figures/fig_7_p6/crops/bbox_14_akimoto_et_al.png" ]
Correspondence
[ "Relation", "Correspondence" ]
PaletteNeRF produces better recoloring results compared to Tan et al. and Akimoto et al. on the lego scene
[]
Figure 7. Qualitative comparison with image palette-based recoloring methods. We compare our method with Tan et al. [31] and Akimoto et al. [1]. For each recolored image, we also show the corresponding palettes editing on its left side, and a zoom-in view on its right side. Ours / Tan et al. Ours / Akimoto et al. Ours ...
[]
The recolored bonsai maintains consistent appearance across different viewing...
PaletteNeRF produces better recoloring results compared to Tan et al. and Akimoto et al. on the lego scene
image B produces better recoloring results compared to Tan et al. and Akimoto et al. on the lego scene
Relation
Correspondence
0.76
paper_0276/fig_7_p6::lego
paper_0269__fig_6_p6__row_6
paper_0269
fig_6_p6
row_6
row_6
Which method demonstrates better highest level of detail in point cloud reconstructions?
[ "(A)", "(B)" ]
(B)
true
[ "LegoFormer", "Ours" ]
[ "paper_0269/figures/fig_6_p6/crops/bbox_18_legoformer.png", "paper_0269/figures/fig_6_p6/crops/bbox_13_ours.png" ]
Detail
[ "Object Form", "Detail" ]
our model produces reconstructions with the highest level of detail
[]
Figure 6. Examples of reconstructions produced by our method along with prior work on seven classes from the ShapeNet-R2N2 dataset [1]. The leftmost image in each row is the conditioning image. Given the simplicity and synthetic nature of ShapeNet, all models produce adequate results; however, our model produces recons...
[]
highest level of detail in point cloud reconstructions
our model produces reconstructions with the highest level of detail
image B model produces reconstructions with the highest level of detail
Object Form
Detail
0.82
paper_0269/fig_6_p6::row_6
paper_0114__fig_4_p6__s3dis
paper_0114
fig_4_p6
s3dis
s3dis
Which method demonstrates better Correct semantic segmentation of point cloud classes with balanced performanc...?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "LwF [21]", "EWC [19]" ]
[ "paper_0114/figures/fig_4_p6/crops/bbox_09_ours.png", "paper_0114/figures/fig_4_p6/crops/bbox_01_lwf.png", "paper_0114/figures/fig_4_p6/crops/bbox_15_ewc.png" ]
Semantic Match
[ "Prompt Match", "Semantic Match" ]
Qualitative comparison of semantic segmentation results on S3DIS and ScanNet datasets, showing that the proposed method achieves better balance between preserving base class knowledge and learning novel classes compared to EWC and LwF forgetting-prevention methods.
[]
Figure 4. Qualitative comparison with the forgetting-prevention methods EWC [19] and LwF [21] on S3DIS and ScanNet datasets of Cnovel = 5 in S0 split. Only the base and novel classes included in current point cloud scenario are explained in the legend. Results in black on ScanNet dataset represent unlabeled and do not ...
[]
Correct semantic segmentation of point cloud classes with balanced performanc...
Qualitative comparison of semantic segmentation results on S3DIS and ScanNet datasets, showing that the proposed method achieves better balance between preserving base class knowledge and learning novel classes compared to EWC and LwF forgetting-prevention methods.
Qualitative comparison of semantic segmentation results on S3DIS and ScanNet datasets, showing that image A achieves better balance between preserving base class knowledge and learning novel classes compared to EWC and LwF forgetting-prevention methods.
Prompt Match
Semantic Match
0.82
paper_0114/fig_4_p6::s3dis
paper_0841__fig_5_p6__scannet
paper_0841
fig_5_p6
scannet
scannet
Which method demonstrates better superior reconstruction of flat surfaces and detailed structures?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "DPT", "MiDaS", "Marigold (ours)", "Omnidata" ]
[ "paper_0841/figures/fig_5_p6/crops/bbox_13_dpt.png", "paper_0841/figures/fig_5_p6/crops/bbox_04_midas.png", "paper_0841/figures/fig_5_p6/crops/bbox_09_marigold.png", "paper_0841/figures/fig_5_p6/crops/bbox_14_omnidata.png" ]
Surface
[ "Object Form", "Surface" ]
Marigold stands out for its superior reconstruction of flat surfaces and detailed structures
[]
Figure 5. Qualitative comparison (unprojected, colored as normals) of monocular depth estimation methods across different datasets. Marigold stands out for its superior reconstruction of flat surfaces and detailed structures. 9498
[]
superior reconstruction of flat surfaces and detailed structures
Marigold stands out for its superior reconstruction of flat surfaces and detailed structures
image C stands out for its superior reconstruction of flat surfaces and detailed structures
Object Form
Surface
0.8
paper_0841/fig_5_p6::scannet
paper_0085__fig_4_p5__scene_3
paper_0085
fig_4_p5
scene_3
scene_3
Which method demonstrates better Camera trajectory accuracy and drift reduction?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "NICE-SLAM", "ESLAM (ours)", "iMAP*" ]
[ "paper_0085/figures/fig_4_p5/crops/bbox_01_nice_slam.png", "paper_0085/figures/fig_4_p5/crops/bbox_00_eslam.png", "paper_0085/figures/fig_4_p5/crops/bbox_02_imap_star.png" ]
Landmark Fidelity
[ "Reference Fidelity", "Landmark Fidelity" ]
ESLAM predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on ScanNet dataset
[]
Figure 4. Qualitative comparison of our proposed ESLAM method’s localization accuracy with existing NeRF-based dense visual SLAM models, iMAP∗[59] and NICE-SLAM [87], on the ScanNet dataset [12]. The ground truth camera trajectory is shown in green, and the estimated trajectory is shown in red. Our method predicts more...
[]
Camera trajectory accuracy and drift reduction
ESLAM predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on ScanNet dataset
image B predicts more accurate camera trajectories and does not suffer from drifting issues compared to iMAP* and NICE-SLAM on ScanNet dataset
Reference Fidelity
Landmark Fidelity
0.82
paper_0085/fig_4_p5::scene_3
paper_0544__fig_6_p5__sample_1
paper_0544
fig_6_p5
sample_1
sample_1
Which method demonstrates better meticulous reconstruction of referenced appearance with side views and 3D car...?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "ours", "Triplanenet", "GOAE", "Zero-1-to-3" ]
[ "paper_0544/figures/fig_6_p5/crops/bbox_06_ours.png", "paper_0544/figures/fig_6_p5/crops/bbox_11_triplanenet.png", "paper_0544/figures/fig_6_p5/crops/bbox_01_goae.png", "paper_0544/figures/fig_6_p5/crops/bbox_08_zero_1_to_3.png" ]
Identity Preservation
[ "Reference Fidelity", "Identity Preservation" ]
DiffPortrait3D demonstrates meticulous reconstruction of referenced appearance, even with side views and 3D cartoon styles, substantially outperforming the baseline methods.
[]
Figure 6. Reconstruction. DiffPortrait3D demonstrates meticu- lous reconstruction of referenced appearance, even with side views and 3D cartoon styles, substantially outperforming the baseline methods. tency module that incorporates cross-view attention within a batch of views. Such a module employ an attention mech- a...
[]
meticulous reconstruction of referenced appearance with side views and 3D car...
DiffPortrait3D demonstrates meticulous reconstruction of referenced appearance, even with side views and 3D cartoon styles, substantially outperforming the baseline methods.
image A demonstrates meticulous reconstruction of referenced appearance, even with side views and 3D cartoon styles, substantially outperforming the baseline methods.
Reference Fidelity
Identity Preservation
0.79
paper_0544/fig_6_p5::sample_1
paper_0687__fig_7_p7__office
paper_0687
fig_7_p7
office
office
Which method demonstrates better Reduced visual artifacts and improved geometry quality in rendered novel views?
[ "(A)", "(B)", "(C)" ]
(C)
true
[ "Pli", "Pli+Sa", "Pli+Sep+Sa" ]
[ "paper_0687/figures/fig_7_p7/crops/bbox_05_pli.png", "paper_0687/figures/fig_7_p7/crops/bbox_01_pli_sa.png", "paper_0687/figures/fig_7_p7/crops/bbox_03_pli_sep_sa.png" ]
Surface
[ "Object Form", "Surface" ]
Ablation study showing progressive improvement in novel view synthesis quality on Blender and LLFF datasets with different component combinations
[]
Figure 7. Qualitative results of ablation studies on Blender and LLFF, where Pli means per-layer inputs incorporation, Sep means separate modeling of colors and volume density, Bkr means back- ground regularization, and Sa means sampling annealing. facts. For forward-facing scenes like LLFF, we find that the sampling an...
[]
Reduced visual artifacts and improved geometry quality in rendered novel views
Ablation study showing progressive improvement in novel view synthesis quality on Blender and LLFF datasets with different component combinations
null
Object Form
Surface
0.82
paper_0687/fig_7_p7::office
paper_1131__fig_5_p4__row_1
paper_1131
fig_5_p4
row_1
row_1
Which method demonstrates better Reduced stretched eyes artifacts and repeated texture artifacts at extrapolat...?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "Our Approach", "Baseline", "Baseline", "Our Approach" ]
[ "paper_1131/figures/fig_5_p4/crops/bbox_02_our_approach_rgb.png", "paper_1131/figures/fig_5_p4/crops/bbox_03_baseline_rgb.png", "paper_1131/figures/fig_5_p4/crops/bbox_14_baseline_normal.png", "paper_1131/figures/fig_5_p4/crops/bbox_04_our_approach_normal.png" ]
Global Artifact
[ "Image Appearance", "Global Artifact" ]
Compared to the baseline, our GMNR better renders the target objects at large non-frontal views due to learning the view-dependent information through the α-VdR module. Stretched eyes can be observed in the case of baseline generated image (row 1, col. 1), while such an artifact is mitigated by our approach (row 1, col...
[]
Figure 5. Synthesized rendered RGB image at a target view (left), its corresponding canonical view (center) along with the mesh (right) for the baseline and our GMNR, respectively. Compared to the baseline, our GMNR better renders the target objects at large non-frontal views due to learning the view-dependent informat...
[]
Reduced stretched eyes artifacts and repeated texture artifacts at extrapolat...
Compared to the baseline, our GMNR better renders the target objects at large non-frontal views due to learning the view-dependent information through the α-VdR module. Stretched eyes can be observed in the case of baseline generated image (row 1, col. 1), while such an artifact is mitigated by our approach (row 1, col...
Compared to the baseline, image A image A better renders the target objects at large non-frontal views due to learning the view-dependent information through the α-VdR module. Stretched eyes can be observed in the case of baseline generated image (row 1, col. 1), while such an artifact is mitigated by image A approach ...
Image Appearance
Global Artifact
0.85
paper_1131/fig_5_p4::row_1
paper_0276__fig_7_p6__fern
paper_0276
fig_7_p6
fern
fern
Which method demonstrates better The recolored bonsai maintains consistent appearance across different viewing...?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "Tan et al.", "Ours", "Akimoto et al." ]
[ "paper_0276/figures/fig_7_p6/crops/bbox_36_tan_et_al.png", "paper_0276/figures/fig_7_p6/crops/bbox_18_ours.png", "paper_0276/figures/fig_7_p6/crops/bbox_06_akimoto_et_al.png" ]
Correspondence
[ "Relation", "Correspondence" ]
PaletteNeRF produces better palette-based recoloring results compared to Tan et al. and Akimoto et al. on the fern scene, with improved view consistency and photorealism
[]
Figure 7. Qualitative comparison with image palette-based recoloring methods. We compare our method with Tan et al. [31] and Akimoto et al. [1]. For each recolored image, we also show the corresponding palettes editing on its left side, and a zoom-in view on its right side. Ours / Tan et al. Ours / Akimoto et al. Ours ...
[]
The recolored bonsai maintains consistent appearance across different viewing...
PaletteNeRF produces better palette-based recoloring results compared to Tan et al. and Akimoto et al. on the fern scene, with improved view consistency and photorealism
image B produces better palette-based recoloring results compared to Tan et al. and Akimoto et al. on the fern scene, with improved view consistency and photorealism
Relation
Correspondence
0.76
paper_0276/fig_7_p6::fern
paper_0108__fig_6_p6__scan77
paper_0108
fig_6_p6
scan77
scan77
Which method demonstrates better accuracy and completeness of depth maps and point cloud reconstruction?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "TransMVSNet", "CasMVSNet", "GeoMVSNet (Ours)", "AA-RMVSNet" ]
[ "paper_0108/figures/fig_6_p6/crops/bbox_00_transmvsnet.png", "paper_0108/figures/fig_6_p6/crops/bbox_06_casmvsnet.png", "paper_0108/figures/fig_6_p6/crops/bbox_15_geomvsnet.png", "paper_0108/figures/fig_6_p6/crops/bbox_09_aa_rmvsnet.png" ]
Completeness
[ "Object Form", "Completeness" ]
GeoMVSNet produces remarkable accuracy and completeness on the most challenging scan77, with significantly accurate depths and complete point clouds, especially for geometry structures, while high-frequency clutter textures are well suppressed
[]
Figure 6. Qualitative comparison of the most challenging scan48 and scan77 on the DTU evaluation dataset. The first and third rows are estimated depth maps while others are point cloud reconstruction results. Our model produces remarkable accuracy and completeness. while the accuracy and completeness of the percentage ...
[]
accuracy and completeness of depth maps and point cloud reconstruction
GeoMVSNet produces remarkable accuracy and completeness on the most challenging scan77, with significantly accurate depths and complete point clouds, especially for geometry structures, while high-frequency clutter textures are well suppressed
image C produces remarkable accuracy and completeness on the most challenging scan77, with significantly accurate depths and complete point clouds, especially for geometry structures, while high-frequency clutter textures are well suppressed
Object Form
Completeness
0.78
paper_0108/fig_6_p6::scan77
paper_1075__fig_3_p4__row_2
paper_1075
fig_3_p4
row_2
row_2
Which method demonstrates better Improved depth map predictions and view synthesis quality from unseen viewpoints?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "DeLiRa (depth and light fields)", "DeLiRa (radiance field)", "Baseline (radiance field)" ]
[ "paper_1075/figures/fig_3_p4/crops/bbox_01_delira_depth_light.png", "paper_1075/figures/fig_3_p4/crops/bbox_03_delira_radiance.png", "paper_1075/figures/fig_3_p4/crops/bbox_05_baseline_radiance.png" ]
Detail
[ "Object Form", "Detail" ]
DeLiRa (depth and light fields) produces better depth and view synthesis results compared to the baseline and DeLiRa (radiance field) variant
[]
Figure 3: Qualitative depth and view synthesis results from unseen viewpoints, using different DeLiRa decoders. As a baseline, we show predictions obtained from a model trained without our contributions, leading to a degenerate learned geometry due to shape-radiance ambiguity (i.e., accurate view synthesis with poor de...
[]
Improved depth map predictions and view synthesis quality from unseen viewpoints
DeLiRa (depth and light fields) produces better depth and view synthesis results compared to the baseline and DeLiRa (radiance field) variant
image A produces better depth and view synthesis results compared to the baseline and DeLiRa (radiance field) variant
Object Form
Detail
0.82
paper_1075/fig_3_p4::row_2
paper_0897__fig_4_p6__buddha_error
paper_0897
fig_4_p6
buddha_error
buddha_error
Which method demonstrates better Fine-grained surface detail recovery and reconstruction accuracy compared to ...?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "MVPSNet [35]", "Ours", "UA-MVPS [12]", "B-MVPS [16]" ]
[ "paper_0897/figures/fig_4_p6/crops/bbox_12_mvpsnet.png", "paper_0897/figures/fig_4_p6/crops/bbox_10_ours.png", "paper_0897/figures/fig_4_p6/crops/bbox_04_ua_mvps.png", "paper_0897/figures/fig_4_p6/crops/bbox_25_b_mvps.png" ]
Geometry Fidelity
[ "Reference Fidelity", "Geometry Fidelity" ]
SuperNormal recovers fine-grained surface details on par with GT meshes in the Buddha object reconstruction, with lower error compared to competing methods
[]
Figure 4. Qualitative comparison of recovered shapes of Buddha and Pot2 and their error maps. Best viewed on screen. Table 2. Quantitative evaluation using normal maps estimated by different methods as inputs. The metrics are averaged over five DiLiGenT-MV [16] objects. Darker colors indicate better results. L2 Chamfer ...
[]
Fine-grained surface detail recovery and reconstruction accuracy compared to ...
SuperNormal recovers fine-grained surface details on par with GT meshes in the Buddha object reconstruction, with lower error compared to competing methods
image B recovers fine-grained surface details on par with GT meshes in the Buddha object reconstruction, with lower error compared to competing methods
Reference Fidelity
Geometry Fidelity
0.93
paper_0897/fig_4_p6::buddha_error
paper_1304__fig_4_p5__zoom
paper_1304
fig_4_p5
zoom
zoom
Which method demonstrates better Reduced fog artifacts and improved scene clarity in rendered outputs?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "Original", "ScatterNSG", "NSG", "ScatterNSG (descattered)" ]
[ "paper_1304/figures/fig_4_p5/crops/bbox_02_original.png", "paper_1304/figures/fig_4_p5/crops/bbox_05_scatter_nsg.png", "paper_1304/figures/fig_4_p5/crops/bbox_01_nsg.png", "paper_1304/figures/fig_4_p5/crops/bbox_04_scatter_nsg_descattered.png" ]
Global Artifact
[ "Image Appearance", "Global Artifact" ]
ScatterNSG produces better visual quality compared to NSG in fog/scattering conditions
[]
Figure 4: Qualitative comparison between Neural Scene Graphs (NSG) [47] and its combination with our proposed network (ScatterNSG). 5.1. Ablation In order to assess the role and contribution of the differ- ent components of our framework, we conduct an ablation study whose results are presented in Tab. 3. We consider
[]
Reduced fog artifacts and improved scene clarity in rendered outputs
ScatterNSG produces better visual quality compared to NSG in fog/scattering conditions
image D produces better visual quality compared to NSG in fog/scattering conditions
Image Appearance
Global Artifact
0.79
paper_1304/fig_4_p5::zoom
paper_0884__fig_3_p7__novel_view_s1
paper_0884
fig_3_p7
novel_view_s1
novel_view_s1
Which method demonstrates better Rendering quality and fidelity of color and depth across novel and training v...?
[ "(A)", "(B)" ]
(B)
true
[ "PS [30]", "Ours" ]
[ "paper_0884/figures/fig_3_p7/crops/bbox_06_ps_30.png", "paper_0884/figures/fig_3_p7/crops/bbox_08_ours.png" ]
Reconstruction Fidelity
[ "Reference Fidelity", "Reconstruction Fidelity" ]
SplaTAM renders color & depth for novel & train views with fidelity comparable to ground truth, while Point-SLAM fails to provide good renderings on both novel & train view images
[]
Figure 3. Renderings on ScanNet++ [49]. Our method, SplaTAM, renders color & depth for the novel & train views with fidelity comparable to the ground truth. It can also be observed that Point-SLAM [30] fails to provide good renderings on both novel & train view images. Note that Point-SLAM (PS) uses the ground-truth de...
[]
Rendering quality and fidelity of color and depth across novel and training v...
SplaTAM renders color & depth for novel & train views with fidelity comparable to ground truth, while Point-SLAM fails to provide good renderings on both novel & train view images
image B renders color & depth for novel & train views with fidelity comparable to ground truth, while Point-SLAM fails to provide good renderings on both novel & train view images
Reference Fidelity
Reconstruction Fidelity
0.88
paper_0884/fig_3_p7::novel_view_s1
paper_0256__fig_3_p4__scan_24
paper_0256
fig_3_p4
scan_24
scan_24
Which method demonstrates better surface accuracy and fidelity?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "NeuS", "NeuralWarp" ]
[ "paper_0256/figures/fig_3_p4/crops/bbox_00_ours.png", "paper_0256/figures/fig_3_p4/crops/bbox_11_neus.png", "paper_0256/figures/fig_3_p4/crops/bbox_14_neural_warp.png" ]
Surface
[ "Object Form", "Surface" ]
Neuralangelo produces more accurate and higher-fidelity surfaces compared to NeuS and NeuralWarp on the DTU benchmark
[]
Figure 3. Qualitative comparison on the DTU benchmark [11]. Neuralangelo produces more accurate and higher-fidelity surfaces. finer hash grids throughout optimization when ϵ decreases to their spatial size. The relearning process can thus be avoided to better capture the details. In practice, we also apply weight decay...
[]
surface accuracy and fidelity
Neuralangelo produces more accurate and higher-fidelity surfaces compared to NeuS and NeuralWarp on the DTU benchmark
image A produces more accurate and higher-fidelity surfaces compared to NeuS and NeuralWarp on the DTU benchmark
Object Form
Surface
0.84
paper_0256/fig_3_p4::scan_24
paper_0114__fig_4_p6__scannet
paper_0114
fig_4_p6
scannet
scannet
Which method demonstrates better Correct semantic segmentation of point cloud classes with balanced performanc...?
[ "(A)", "(B)", "(C)" ]
(C)
true
[ "EWC [19]", "LwF [21]", "Ours" ]
[ "paper_0114/figures/fig_4_p6/crops/bbox_00_ewc.png", "paper_0114/figures/fig_4_p6/crops/bbox_13_lwf.png", "paper_0114/figures/fig_4_p6/crops/bbox_03_ours.png" ]
Semantic Match
[ "Prompt Match", "Semantic Match" ]
Qualitative comparison of semantic segmentation on ScanNet dataset showing that the proposed method achieves better balance between preserving base class knowledge and learning novel classes compared to EWC and LwF forgetting-prevention methods
[]
Figure 4. Qualitative comparison with the forgetting-prevention methods EWC [19] and LwF [21] on S3DIS and ScanNet datasets of Cnovel = 5 in S0 split. Only the base and novel classes included in current point cloud scenario are explained in the legend. Results in black on ScanNet dataset represent unlabeled and do not ...
[]
Correct semantic segmentation of point cloud classes with balanced performanc...
Qualitative comparison of semantic segmentation on ScanNet dataset showing that the proposed method achieves better balance between preserving base class knowledge and learning novel classes compared to EWC and LwF forgetting-prevention methods
Qualitative comparison of semantic segmentation on ScanNet dataset showing that image C achieves better balance between preserving base class knowledge and learning novel classes compared to EWC and LwF forgetting-prevention methods
Prompt Match
Semantic Match
0.82
paper_0114/fig_4_p6::scannet
paper_0337__fig_6_p6__hand_zoom
paper_0337
fig_6_p6
hand_zoom
hand_zoom
Which method demonstrates better hand articulation and facial expression accuracy?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "SCANimate", "SNARF" ]
[ "paper_0337/figures/fig_6_p6/crops/bbox_17_ours.png", "paper_0337/figures/fig_6_p6/crops/bbox_10_scanimate.png", "paper_0337/figures/fig_6_p6/crops/bbox_09_snarf.png" ]
Anatomy
[ "Object Form", "Anatomy" ]
Our method recovers hand articulation and facial expression most accurately
[]
Figure 6. Qualitative results on GRAB dataset. Our method recovers hand articulation and facial expression most accurately. 4.4. Results on GRAB Dataset Tab. 2 summarizes results on the GRAB dataset. Over- all, our method beats all baselines, especially for the hands, where the margin is large. Fig. 6 visually shows th...
[]
hand articulation and facial expression accuracy
Our method recovers hand articulation and facial expression most accurately
image A method recovers hand articulation and facial expression most accurately
Object Form
Anatomy
0.82
paper_0337/fig_6_p6::hand_zoom
paper_0139__fig_1_p0__table
paper_0139
fig_1_p0
table
table
Which method demonstrates better Correct instance segmentation and separation of nearby chair instances withou...?
[ "(A)", "(B)" ]
(B)
true
[ "DyCo3D's", "Ours" ]
[ "paper_0139/figures/fig_1_p0/crops/bbox_09_dyco3ds.png", "paper_0139/figures/fig_1_p0/crops/bbox_02_ours.png" ]
Correspondence
[ "Relation", "Correspondence" ]
ISBNet addresses clustering issues in DyCo3D by achieving far better instance segmentation results through instance-aware point sampling
[]
Figure 1. In DyCo3D [16], kernel prediction quality is greatly affected by the centroid-based clustering algorithm which has two issues: 1 mis-grouping nearby instances and 2 over-segment a large object into multiple fragments. Our method addresses these issues by instance-aware point sampling, achieving far better res...
[]
Correct instance segmentation and separation of nearby chair instances withou...
ISBNet addresses clustering issues in DyCo3D by achieving far better instance segmentation results through instance-aware point sampling
image B addresses clustering issues in DyCo3D by achieving far better instance segmentation results through instance-aware point sampling
Relation
Correspondence
0.78
paper_0139/fig_1_p0::table
paper_0842__fig_5_p7__row_2
paper_0842
fig_5_p7
row_2
row_2
Which method demonstrates better Segmentation accuracy and correctness of point cloud semantic labels?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "Support Mask", "Query", "w/ BPC", "w/o BPC" ]
[ "paper_0842/figures/fig_5_p7/crops/bbox_03_support_mask.png", "paper_0842/figures/fig_5_p7/crops/bbox_09_query.png", "paper_0842/figures/fig_5_p7/crops/bbox_01_w_bpc.png", "paper_0842/figures/fig_5_p7/crops/bbox_02_wo_bpc.png" ]
Attribute Binding
[ "Relation", "Attribute Binding" ]
Visual comparison of segmentation results with BPC (w/ BPC) and without BPC (w/o BPC) on 1-way 1-shot tasks for bookcase and chair classes
[]
Figure 5. Visual comparisons between our models with BPC (w/ BPC) and without BPC (w/o BPC). Each row corresponds to the 1-way 1-shot task targeting bookcase (green) and chair (red), respectively, arranged from top to bottom. 5.3. Ablation Study In this section, we report the mIoU results as the mean of all splits of S...
[]
Segmentation accuracy and correctness of point cloud semantic labels
Visual comparison of segmentation results with BPC (w/ BPC) and without BPC (w/o BPC) on 1-way 1-shot tasks for bookcase and chair classes
Visual comparison of segmentation results with BPC (image C) and without BPC (w/o BPC) on 1-way 1-shot tasks for bookcase and chair classes
Relation
Attribute Binding
0.75
paper_0842/fig_5_p7::row_2
paper_0093__fig_9_p7__stair
paper_0093
fig_9_p7
stair
stair
Which method demonstrates better image clarity?
[ "(A)", "(B)" ]
(A)
true
[ "Pers. warp", "Inv. warp" ]
[ "paper_0093/figures/fig_9_p7/crops/bbox_11_pers_warp.png", "paper_0093/figures/fig_9_p7/crops/bbox_00_inv_warp.png" ]
Blur
[ "Image Appearance", "Blur" ]
Our perspective warping renders more clear images than inverse sphere warping
[]
Figure 9. Visual comparison of different warping techniques. Our perspective warping renders more clear images than inverse sphere warping [3]. Setting PSNR↑ SSIM↑ LPIPS↓
[]
image clarity
Our perspective warping renders more clear images than inverse sphere warping
image A perspective warping renders more clear images than inverse sphere warping
Image Appearance
Blur
0.86
paper_0093/fig_9_p7::stair
paper_1237__fig_6_p6__roughness
paper_1237
fig_6_p6
roughness
roughness
Which method demonstrates better Smoother normal estimation, more accurate environment map, better relighting ...?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "InvRender", "NeRFactor" ]
[ "paper_1237/figures/fig_6_p6/crops/bbox_11_ours.png", "paper_1237/figures/fig_6_p6/crops/bbox_03_invrender.png", "paper_1237/figures/fig_6_p6/crops/bbox_14_nerfactor.png" ]
Surface
[ "Object Form", "Surface" ]
NeMF produces better roughness estimation compared to NeRFactor and InvRender, with smoother and more accurate results closer to ground truth
[]
Figure 6: Comparisons with other approaches. We show the normal, albedo, roughness, and environment map estimated by NeRFactor, InvRender and our method on three scenes. We also compared the rendering results of the new viewpoint under three different lighting conditions (the original light and two novel lights). Note ...
[]
Smoother normal estimation, more accurate environment map, better relighting ...
NeMF produces better roughness estimation compared to NeRFactor and InvRender, with smoother and more accurate results closer to ground truth
image A produces better roughness estimation compared to NeRFactor and InvRender, with smoother and more accurate results closer to ground truth
Object Form
Surface
0.78
paper_1237/fig_6_p6::roughness
paper_0781__fig_9_p7__main_sample
paper_0781
fig_9_p7
main_sample
main_sample
Which method demonstrates better Novel-view synthesis quality and completeness?
[ "(A)", "(B)", "(C)" ]
(B)
true
[ "Zero* I-to-3", "Ours", "Zero I-to-3" ]
[ "paper_0781/figures/fig_9_p7/crops/bbox_04_zero_star_i_to_3.png", "paper_0781/figures/fig_9_p7/crops/bbox_00_ours.png", "paper_0781/figures/fig_9_p7/crops/bbox_02_zero_i_to_3.png" ]
Action Match
[ "Prompt Match", "Action Match" ]
None result in good novel-view synthesis, except NViST
[]
Figure 9. Qualitative comparison with Zero-1-to-3 [48] on a phone captured input image: Zero-1-to-3 shows result w/masked input, Zero1to3∗shows result w/full image input. None result in good novel-view synthesis, except NViST. PSNR↑ SSIM↑
[]
Novel-view synthesis quality and completeness
None result in good novel-view synthesis, except NViST
None result in good novel-view synthesis, except image B
Prompt Match
Action Match
0.86
paper_0781/fig_9_p7::main_sample
paper_1243__fig_6_p7__row_1
paper_1243
fig_6_p7
row_1
row_1
Which method demonstrates better Restoration of bent lines in underwater scene?
[ "(A)", "(B)" ]
(A)
true
[ "water-free(NeRFrac)", "water-free(LB-NeRF)" ]
[ "paper_1243/figures/fig_6_p7/crops/bbox_07_water_free_nerfrac.png", "paper_1243/figures/fig_6_p7/crops/bbox_00_water_free_lb_nerf.png" ]
Detail
[ "Object Form", "Detail" ]
NeRFrac successfully restores bent lines thanks to the explicit modeling of Snell's Law, while LB-NeRF fails in restoring water-free images
[]
Figure 6: Qualitative result of underwater scene restoration. We can render water-free scenes by removing the Refrac- tive Field. We directly remove the deformation field of LB- NeRF[13] to serve as a comparison. As can be seen, our method successfully restores bent lines thanks to the explicit modeling of Snell’s Law,...
[]
Restoration of bent lines in underwater scene
NeRFrac successfully restores bent lines thanks to the explicit modeling of Snell's Law, while LB-NeRF fails in restoring water-free images
image A successfully restores bent lines thanks to the explicit modeling of Snell's Law, while LB-NeRF fails in restoring water-free images
Object Form
Detail
0.72
paper_1243/fig_6_p7::row_1
paper_0639__fig_1_p0__precision
paper_0639
fig_1_p0
precision
precision
Which method demonstrates better Precision and recall error maps showing reconstruction accuracy?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "(e) Ours", "(c) ET-MVSNet [14]", "(d) GeoMVSNet [46]", "(b) RA-MVSNet [45]" ]
[ "paper_0639/figures/fig_1_p0/crops/bbox_00_ours.png", "paper_0639/figures/fig_1_p0/crops/bbox_09_et_mvsnet.png", "paper_0639/figures/fig_1_p0/crops/bbox_10_geomvsnet.png", "paper_0639/figures/fig_1_p0/crops/bbox_06_ra_mvsnet.png" ]
Geometry Fidelity
[ "Reference Fidelity", "Geometry Fidelity" ]
GoMVS demonstrates notable improvements over existing methods in challenging areas on the Tanks and Temple benchmark Horse scan, as shown by precision and recall error maps.
[]
Figure 1. Comparison of reconstruction errors on Tanks and Temple benchmark. We show precision and recall error maps for the “Horse” scan. Our method demonstrates notable improvements over existing methods in challenging areas. Abstract Matching cost aggregation plays a fundamental role in learning-based multi-view ste...
[]
Precision and recall error maps showing reconstruction accuracy
GoMVS demonstrates notable improvements over existing methods in challenging areas on the Tanks and Temple benchmark Horse scan, as shown by precision and recall error maps.
image A demonstrates notable improvements over existing methods in challenging areas on the Tanks and Temple benchmark Horse scan, as shown by precision and recall error maps.
Reference Fidelity
Geometry Fidelity
0.79
paper_0639/fig_1_p0::precision
paper_1287__fig_2_p1__scene_c
paper_1287
fig_2_p1
scene_c
scene_c
Which method demonstrates better Correct semantic segmentation of point cloud objects with improved distinctio...?
[ "(A)", "(B)" ]
(A)
true
[ "Retro-FPN", "Point Transformer" ]
[ "paper_1287/figures/fig_2_p1/crops/bbox_04_retro_fpn.png", "paper_1287/figures/fig_2_p1/crops/bbox_03_point_transformer.png" ]
Correspondence
[ "Relation", "Correspondence" ]
Retro-FPN improves semantic segmentation by refining intermediate predictions, recovering lost information (column) and better distinguishing between similar objects (chair vs sofa) compared to Point Transformer backbone
[]
Figure 2. Visualization of segmentation process of Retro-FPN. (a) and (c) show the visual comparison with the backbone (Point Transformer [87]) network. In (a), the backbone loses the infor- mation of the column. In (c), the backbone struggles to distinguish between chair and sofa. In (b) and (d), we show the retrospec...
[]
Correct semantic segmentation of point cloud objects with improved distinctio...
Retro-FPN improves semantic segmentation by refining intermediate predictions, recovering lost information (column) and better distinguishing between similar objects (chair vs sofa) compared to Point Transformer backbone
image A improves semantic segmentation by refining intermediate predictions, recovering lost information (column) and better distinguishing between similar objects (chair vs sofa) compared to Point Transformer backbone
Relation
Correspondence
0.72
paper_1287/fig_2_p1::scene_c
paper_1263__fig_7_p7__row_5
paper_1263
fig_7_p7
row_5
row_5
Which method demonstrates better Ability to reconstruct 3D models from incomplete line drawings with only visi...?
[ "(A)", "(B)" ]
(B)
true
[ "Front", "Ours" ]
[ "paper_1263/figures/fig_7_p7/crops/bbox_44_front.png", "paper_1263/figures/fig_7_p7/crops/bbox_18_ours.png" ]
Completeness
[ "Object Form", "Completeness" ]
PlankAssembly correctly reconstructs objects from noisy line drawings (10%, 20%, 30% noise levels) and incomplete inputs with visible parts only, while traditional pipeline fails or degrades significantly
[]
Figure 7. Qualitative results. Rows 1-2: clean inputs. Rows 3-5: Noisy inputs with noise level 10%, 20%, and 30%, respectively. Row 6: Inputs with visible parts only. We use red boxes to indicate incorrect reconstructions. 0 10 20 30 Noise Level (%)
[]
Ability to reconstruct 3D models from incomplete line drawings with only visi...
PlankAssembly correctly reconstructs objects from noisy line drawings (10%, 20%, 30% noise levels) and incomplete inputs with visible parts only, while traditional pipeline fails or degrades significantly
image B correctly reconstructs objects from noisy line drawings (10%, 20%, 30% noise levels) and incomplete inputs with visible parts only, while traditional pipeline fails or degrades significantly
Object Form
Completeness
0.78
paper_1263/fig_7_p7::row_5
paper_1235__fig_4_p6__row_3
paper_1235
fig_4_p6
row_3
row_3
Which method demonstrates better detail level and completeness of reconstructed avatar geometry?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "Ours", "DSFN", "SMPL", "PINA" ]
[ "paper_1235/figures/fig_4_p6/crops/bbox_05_ours.png", "paper_1235/figures/fig_4_p6/crops/bbox_07_dsfn.png", "paper_1235/figures/fig_4_p6/crops/bbox_09_smpl.png", "paper_1235/figures/fig_4_p6/crops/bbox_08_pina.png" ]
Detail
[ "Object Form", "Detail" ]
Our method reconstructs more detailed avatars compared to SMPL, SMPL+D, DSFN, and PINA on partial point cloud reconstruction task
[]
Figure 4: Partial point cloud reconstruction on BuFF [69]: We first compare with fitting SMPL and SMPL+D models to our partial point clouds and then compare against more contemporary baselines DSFN [6] and PINA [13]. Our method reconstructs more detailed avatars. reported in Tab. 1 and Fig. 4, 5. The results for each i...
[]
detail level and completeness of reconstructed avatar geometry
Our method reconstructs more detailed avatars compared to SMPL, SMPL+D, DSFN, and PINA on partial point cloud reconstruction task
image A method reconstructs more detailed avatars compared to SMPL, SMPL+D, DSFN, and PINA on partial point cloud reconstruction task
Object Form
Detail
0.79
paper_1235/fig_4_p6::row_3
paper_1141__fig_6_p6__row_2
paper_1141
fig_6_p6
row_2
row_2
Which method demonstrates better hair direction preservation?
[ "(A)", "(B)", "(C)", "(D)" ]
(B)
true
[ "Identity Hair", "HairNeRF", "Barbershop", "StyY.Hair" ]
[ "paper_1141/figures/fig_6_p6/crops/bbox_05_identity_hair.png", "paper_1141/figures/fig_6_p6/crops/bbox_13_hairnerf.png", "paper_1141/figures/fig_6_p6/crops/bbox_04_barbershop.png", "paper_1141/figures/fig_6_p6/crops/bbox_14_styy_hair.png" ]
Shape
[ "Object Form", "Shape" ]
our method preserves the direction of the hair better
[]
Figure 6: Comparison of results for FFHQ datasets. Both StyleYourHair and our method align the hair even if the poses between the input images are different, but our method preserves the direction of the hair better (first and the last rows). ="GqiHwyxBQ7t2exuiSmBi8xrTEs=">AC1H icjVHLSsNAFD2Nr1ofjbp0EyCq5KIqMuiG91Vs A9...
[]
hair direction preservation
our method preserves the direction of the hair better
image B method preserves the direction of the hair better
Object Form
Shape
0.82
paper_1141/fig_6_p6::row_2
paper_0276__fig_9_p7__horns
paper_0276
fig_9_p7
horns
horns
Which method demonstrates better spatially and temporally consistent style transfer results?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "Style Image", "ARF", "Ours", "AdaIN" ]
[ "paper_0276/figures/fig_9_p7/crops/bbox_14_style_image.png", "paper_0276/figures/fig_9_p7/crops/bbox_05_arf.png", "paper_0276/figures/fig_9_p7/crops/bbox_00_ours.png", "paper_0276/figures/fig_9_p7/crops/bbox_09_adain.png" ]
Correspondence
[ "Relation", "Correspondence" ]
Our method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism.
[]
Figure 9. Qualitative comparison with baseline stylization methods. Our method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism. Reference w/o ℒ!" w/o ℒ!# w/o 𝜹 Ours
[]
spatially and temporally consistent style transfer results
Our method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism.
image C method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism.
Relation
Correspondence
0.74
paper_0276/fig_9_p7::horns
paper_0102__fig_5_p5__materials
paper_0102
fig_5_p5
materials
materials
Which method demonstrates better sharper detailed textures without artifacts?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "CCNeRF", "NeRF", "Huang et al.", "Ours" ]
[ "paper_0102/figures/fig_5_p5/crops/bbox_18_ccnerf.png", "paper_0102/figures/fig_5_p5/crops/bbox_15_nerf.png", "paper_0102/figures/fig_5_p5/crops/bbox_22_huang_et_al.png", "paper_0102/figures/fig_5_p5/crops/bbox_01_ours.png" ]
Texture
[ "Object Form", "Texture" ]
Our method renders sharper detailed textures without artifacts.
[]
Figure 5. Qualitative evaluations on NeRF-Synthetic [21] dataset. Our method renders sharper detailed textures without artifacts. shown in Fig. 4, we first rasterize the combined point clouds to determine which scene each pixel belongs to and gener- ate a mask for each scene. Then we render each scene sepa- rately and ...
[]
sharper detailed textures without artifacts
Our method renders sharper detailed textures without artifacts.
image D method renders sharper detailed textures without artifacts.
Object Form
Texture
0.88
paper_0102/fig_5_p5::materials
paper_0276__fig_9_p7__orchids
paper_0276
fig_9_p7
orchids
orchids
Which method demonstrates better spatially and temporally consistent style transfer results?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "Ours", "UPST-NeRF", "AdaIN", "ARF" ]
[ "paper_0276/figures/fig_9_p7/crops/bbox_18_ours.png", "paper_0276/figures/fig_9_p7/crops/bbox_02_upst_nerf.png", "paper_0276/figures/fig_9_p7/crops/bbox_17_adain.png", "paper_0276/figures/fig_9_p7/crops/bbox_13_arf.png" ]
Correspondence
[ "Relation", "Correspondence" ]
Our method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism.
[]
Figure 9. Qualitative comparison with baseline stylization methods. Our method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism. Reference w/o ℒ!" w/o ℒ!# w/o 𝜹 Ours
[]
spatially and temporally consistent style transfer results
Our method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism.
image A method produces spatially and temporally consistent style transfer results that faithfully match the style examples while maintaining photorealism.
Relation
Correspondence
0.74
paper_0276/fig_9_p7::orchids
paper_0011__fig_7_p6__row_1
paper_0011
fig_7_p6
row_1
row_1
Which method demonstrates better ALTO achieves detailed qualitative results with higher fidelity reconstructio...?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "(d) ConvONet [52]", "(c) SPSR [37]", "(f) ALTO", "(e) POCO [4]" ]
[ "paper_0011/figures/fig_7_p6/crops/bbox_13_convonet.png", "paper_0011/figures/fig_7_p6/crops/bbox_12_spsr.png", "paper_0011/figures/fig_7_p6/crops/bbox_02_alto.png", "paper_0011/figures/fig_7_p6/crops/bbox_15_poco.png" ]
Detail
[ "Object Form", "Detail" ]
ALTO achieves detailed qualitative results with higher fidelity reconstruction, better preserving fine details like the double-deck table and chair slats compared to ConvONet and POCO baselines
[]
Figure 7. Qualitative comparison on scene-level reconstruction Synthetic Room Dataset. Learning-based methods are trained and tested on 10K noisy points. ALTO can reconstruct the (top scene) double-deck table and (bottom scene) details in the chair. Method IoU ↑ Chamfer-L1 ↓ NC↑ F-score↑
[]
ALTO achieves detailed qualitative results with higher fidelity reconstructio...
ALTO achieves detailed qualitative results with higher fidelity reconstruction, better preserving fine details like the double-deck table and chair slats compared to ConvONet and POCO baselines
image C achieves detailed qualitative results with higher fidelity reconstruction, better preserving fine details like the double-deck table and chair slats compared to ConvONet and POCO baselines
Object Form
Detail
0.81
paper_0011/fig_7_p6::row_1
paper_0074__fig_1_p0__scene_2
paper_0074
fig_1_p0
scene_2
scene_2
Which method demonstrates better clearer edges and more accurate geometry?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "NeRF", "ADfSM", "Panorama", "Ours" ]
[ "paper_0074/figures/fig_1_p0/crops/bbox_02_nerf.png", "paper_0074/figures/fig_1_p0/crops/bbox_01_adfsm.png", "paper_0074/figures/fig_1_p0/crops/bbox_04_panorama.png", "paper_0074/figures/fig_1_p0/crops/bbox_03_ours.png" ]
Boundary
[ "Object Form", "Boundary" ]
Compared with state-of-the-art methods, measurements from ours have clearer edges and more accurate geometry.
[]
Figure 1. Our work aims to implicitly represent a scene with a neural network through fewer panoramas and thus obtain the relative depth measurements. The left diagram illustrates the efficiency and accuracy of our method against NeRF [18] and a traditional geometric-based Structure from Motion approach [12]. All metho...
[]
clearer edges and more accurate geometry
Compared with state-of-the-art methods, measurements from ours have clearer edges and more accurate geometry.
Compared with state-of-the-art methods, measurements from image D have clearer edges and more accurate geometry.
Object Form
Boundary
0.78
paper_0074/fig_1_p0::scene_2
paper_0326__fig_9_p7__row_3
paper_0326
fig_9_p7
row_3
row_3
Which method demonstrates better Plane segmentation accuracy in edge regions?
[ "(A)", "(B)" ]
(A)
true
[ "Ours", "Plane R-CNN" ]
[ "paper_0326/figures/fig_9_p7/crops/bbox_10_ours.png", "paper_0326/figures/fig_9_p7/crops/bbox_01_plane_r_cnn.png" ]
Segmentation
[ "Object Form", "Segmentation" ]
Our method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN
[]
Figure 9. Plane Segmentation Comparisons. Our method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN [18]. Plane Reconstruction. To demonstrate the capability of our method on the unsupervised plane reconstruction task, we compare its reconstruction performance with the sta...
[]
Plane segmentation accuracy in edge regions
Our method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN
image A method can generate way more accurate segmentation results in edge regions compared with Plane R-CNN
Object Form
Segmentation
0.75
paper_0326/fig_9_p7::row_3
paper_0226__fig_1_p0__row_2
paper_0226
fig_1_p0
row_2
row_2
Which method demonstrates better Better geometry estimations and reconstruction accuracy compared to COLMAP, N...?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "PatchmatchNet", "ACMMP", "NeuS", "Ours" ]
[ "paper_0226/figures/fig_1_p0/crops/bbox_13_patchmatchnet.png", "paper_0226/figures/fig_1_p0/crops/bbox_03_acmmp.png", "paper_0226/figures/fig_1_p0/crops/bbox_16_neus.png", "paper_0226/figures/fig_1_p0/crops/bbox_15_ours.png" ]
Shape
[ "Object Form", "Shape" ]
Our approach outperforms existing approaches over standard 3D benchmarks with better geometry estimations
[]
Figure 1. Reconstructions with various methods using 14 images of a model from BlendedMVS [70]. locations in 3D where 2D image observations concur, a property called photo-consistency. This observation con- sistency strategy has been later challenged by approaches in the field that seek instead for observation fidelity u...
[]
Better geometry estimations and reconstruction accuracy compared to COLMAP, N...
Our approach outperforms existing approaches over standard 3D benchmarks with better geometry estimations
image D approach outperforms existing approaches over standard 3D benchmarks with better geometry estimations
Object Form
Shape
0.81
paper_0226/fig_1_p0::row_2
paper_1294__fig_1_p0__row_1
paper_1294
fig_1_p0
row_1
row_1
Which method demonstrates better fewer visual artifacts, geometric deformities, and blurs?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "MPI", "MINE", "Ours", "PixelSynth" ]
[ "paper_1294/figures/fig_1_p0/crops/bbox_16_mpi.png", "paper_1294/figures/fig_1_p0/crops/bbox_11_mine.png", "paper_1294/figures/fig_1_p0/crops/bbox_14_ours.png", "paper_1294/figures/fig_1_p0/crops/bbox_09_pixelsynth.png" ]
Blur
[ "Image Appearance", "Blur" ]
our method synthesizes novel views with fewer visual artifacts, geometric deformities, and blurs. Notably, our method models favorable intricate details, such as tiny objects, symbols, and traffic signs, resulting in more photo-realistic views.
[]
Figure 1: Novel view synthesis result comparisons. Given a single image captured in an outdoor scene, our method synthesizes novel views with fewer visual artifacts, geometric deformities, and blurs. Notably, our method models favorable intricate details, such as tiny objects, symbols, and traffic signs, resulting in m...
[]
fewer visual artifacts, geometric deformities, and blurs
our method synthesizes novel views with fewer visual artifacts, geometric deformities, and blurs. Notably, our method models favorable intricate details, such as tiny objects, symbols, and traffic signs, resulting in more photo-realistic views.
image C method synthesizes novel views with fewer visual artifacts, geometric deformities, and blurs. Notably, image C method models favorable intricate details, such as tiny objects, symbols, and traffic signs, resulting in more photo-realistic views.
Image Appearance
Blur
0.72
paper_1294/fig_1_p0::row_1
paper_0968__fig_3_p4__row_5
paper_0968
fig_3_p4
row_5
row_5
Which method demonstrates better Darker error maps indicating better depth estimation uniformly across image?
[ "(A)", "(B)", "(C)", "(D)" ]
(C)
true
[ "Adabins", "Adabins Error", "Ours Error", "Image" ]
[ "paper_0968/figures/fig_3_p4/crops/bbox_24_adabins.png", "paper_0968/figures/fig_3_p4/crops/bbox_14_adabins_error.png", "paper_0968/figures/fig_3_p4/crops/bbox_04_ours_error.png", "paper_0968/figures/fig_3_p4/crops/bbox_12_image.png" ]
Reconstruction Fidelity
[ "Reference Fidelity", "Reconstruction Fidelity" ]
WorDepth improves uniformly across the image (darker in error map), implying better scale. WorDepth also predicts more accurate depth in regions corresponding to chairs, window, shower curtain, man, and desks, which are objects specified by text descriptions.
[]
Figure 3. Qualitative results on NYU Depth V2. We compare WorDepth with AdaBins [2]. Text descriptions are generated using ExpansionNet v2 [25]. Overall, WorDepth improves uniformly across the image (darker in error map), implying better scale. WorDepth also predicts more accurate depth in regions corresponding to “cha...
[]
Darker error maps indicating better depth estimation uniformly across image
WorDepth improves uniformly across the image (darker in error map), implying better scale. WorDepth also predicts more accurate depth in regions corresponding to chairs, window, shower curtain, man, and desks, which are objects specified by text descriptions.
image C improves uniformly across the image (darker in error map), implying better scale. image C also predicts more accurate depth in regions corresponding to chairs, window, shower curtain, man, and desks, which are objects specified by text descriptions.
Reference Fidelity
Reconstruction Fidelity
0.85
paper_0968/fig_3_p4::row_5
paper_0276__fig_7_p6__bonsai
paper_0276
fig_7_p6
bonsai
bonsai
Which method demonstrates better The recolored bonsai maintains consistent appearance across different viewing...?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "Ours", "Akimoto et al.", "Tan et al." ]
[ "paper_0276/figures/fig_7_p6/crops/bbox_24_ours.png", "paper_0276/figures/fig_7_p6/crops/bbox_02_akimoto_et_al.png", "paper_0276/figures/fig_7_p6/crops/bbox_01_tan_et_al.png" ]
Correspondence
[ "Relation", "Correspondence" ]
PaletteNeRF produces better qualitative recoloring results compared to Tan et al. and Akimoto et al. on the bonsai scene
[]
Figure 7. Qualitative comparison with image palette-based recoloring methods. We compare our method with Tan et al. [31] and Akimoto et al. [1]. For each recolored image, we also show the corresponding palettes editing on its left side, and a zoom-in view on its right side. Ours / Tan et al. Ours / Akimoto et al. Ours ...
[]
The recolored bonsai maintains consistent appearance across different viewing...
PaletteNeRF produces better qualitative recoloring results compared to Tan et al. and Akimoto et al. on the bonsai scene
image A produces better qualitative recoloring results compared to Tan et al. and Akimoto et al. on the bonsai scene
Relation
Correspondence
0.76
paper_0276/fig_7_p6::bonsai
paper_0246__fig_4_p5__subject_3
paper_0246
fig_4_p5
subject_3
subject_3
Which method demonstrates better richer skin details and high-fidelity highlights?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "NeuFace", "DIFFREC", "PhySG", "VolSDF" ]
[ "paper_0246/figures/fig_4_p5/crops/bbox_05_neuface.png", "paper_0246/figures/fig_4_p5/crops/bbox_01_diffrec.png", "paper_0246/figures/fig_4_p5/crops/bbox_09_physg.png", "paper_0246/figures/fig_4_p5/crops/bbox_16_volsdf.png" ]
Texture
[ "Object Form", "Texture" ]
NeuFace captures faces with much richer skin details and high-fidelity highlights compared to Ref-NeRF, PhySG, VolSDF, and DIFFREC in novel view synthesis
[]
Figure 4. Comparison of novel view synthesis with Ref-NeRF [47], PhySG [56], VolSDF [54] and DIFFREC [31]. Zoom in for a better view. NeuFace captures faces with much richer skin details and high-fidelity highlights. w/ our calibration F. S.N. ↓ Subject 1 (Rich Reflection) Subject 2 (Moderate Reflection)
[]
richer skin details and high-fidelity highlights
NeuFace captures faces with much richer skin details and high-fidelity highlights compared to Ref-NeRF, PhySG, VolSDF, and DIFFREC in novel view synthesis
image A captures faces with much richer skin details and high-fidelity highlights compared to Ref-NeRF, PhySG, VolSDF, and DIFFREC in novel view synthesis
Object Form
Texture
0.79
paper_0246/fig_4_p5::subject_3
paper_0680__fig_2_p1__cat_and_dog
paper_0680
fig_2_p1
cat_and_dog
cat_and_dog
Which method demonstrates better Accurate alignment between generated image and text prompt?
[ "(A)", "(B)", "(C)" ]
(A)
true
[ "INITNO (Ours)", "Stable Diffusion (Valid)", "Stable Diffusion (Invalid)" ]
[ "paper_0680/figures/fig_2_p1/crops/bbox_01_initno_ours.png", "paper_0680/figures/fig_2_p1/crops/bbox_00_stable_diffusion_valid.png", "paper_0680/figures/fig_2_p1/crops/bbox_07_stable_diffusion_invalid.png" ]
Semantic Match
[ "Prompt Match", "Semantic Match" ]
InitNO guides initial noise into valid regions to synthesize high-fidelity results that precisely correspond to text prompts, outperforming Stable Diffusion with invalid noise configurations.
[]
Figure 2. INITNO. Our investigation dives into the exploration of various random noise configurations and their subsequent influence on the generated results. Notably, when different noises are input into SD under identical text prompts, there are marked discrepancy in the alignment between the generated image and the ...
[]
Accurate alignment between generated image and text prompt
InitNO guides initial noise into valid regions to synthesize high-fidelity results that precisely correspond to text prompts, outperforming Stable Diffusion with invalid noise configurations.
image A guides initial noise into valid regions to synthesize high-fidelity results that precisely correspond to text prompts, outperforming Stable Diffusion with invalid noise configurations.
Prompt Match
Semantic Match
0.92
paper_0680/fig_2_p1::cat_and_dog
paper_1276__fig_3_p5__row_4
paper_1276
fig_3_p5
row_4
row_4
Which method demonstrates better Error magnitude visualization where blue indicates lower error, red indicates...?
[ "(A)", "(B)", "(C)", "(D)" ]
(A)
true
[ "RPEFlow (Ours)", "CamLiFlow [13]", "CamLiFlow+Events", "Point Clouds" ]
[ "paper_1276/figures/fig_3_p5/crops/bbox_10_rpeflow.png", "paper_1276/figures/fig_3_p5/crops/bbox_17_camliflow.png", "paper_1276/figures/fig_3_p5/crops/bbox_25_camliflow_events.png", "paper_1276/figures/fig_3_p5/crops/bbox_05_point_clouds.png" ]
Reconstruction Fidelity
[ "Reference Fidelity", "Reconstruction Fidelity" ]
RPEFlow achieves lower scene flow estimation error compared to CamLiFlow and CamLiFlow+Events, as shown by the error visualization in Figure 3 row 4 where blue (lower error) is more prominent in the RPEFlow result.
[]
Figure 3: Visual comparisons on simulated data, in which the top row is on the “val” split of FlyingThings3D [28] dataset and the reset two are on the test split of our simulated EKubric dataset. For the bottom 3D comparisons, blue indicates a lower error, red indicates a higher error, and green indicates the median. B...
[]
Error magnitude visualization where blue indicates lower error, red indicates...
RPEFlow achieves lower scene flow estimation error compared to CamLiFlow and CamLiFlow+Events, as shown by the error visualization in Figure 3 row 4 where blue (lower error) is more prominent in the RPEFlow result.
image A achieves lower scene flow estimation error compared to CamLiFlow and CamLiFlow+Events, as shown by the error visualization in Figure 3 row 4 where blue (lower error) is more prominent in the image A result.
Reference Fidelity
Reconstruction Fidelity
0.8
paper_1276/fig_3_p5::row_4
paper_0011__fig_7_p6__row_3
paper_0011
fig_7_p6
row_3
row_3
Which method demonstrates better ALTO achieves detailed qualitative results with higher fidelity reconstructio...?
[ "(A)", "(B)", "(C)", "(D)" ]
(D)
true
[ "(e) POCO [4]", "SPSR [37]", "(d) ConvONet [52]", "(f) ALTO" ]
[ "paper_0011/figures/fig_7_p6/crops/bbox_07_poco.png", "paper_0011/figures/fig_7_p6/crops/bbox_00_spsr.png", "paper_0011/figures/fig_7_p6/crops/bbox_14_convonet.png", "paper_0011/figures/fig_7_p6/crops/bbox_03_alto.png" ]
Detail
[ "Object Form", "Detail" ]
ALTO achieves detailed qualitative results compared to baselines, reconstructing fine details such as the double-deck table and chair details with higher fidelity
[]
Figure 7. Qualitative comparison on scene-level reconstruction Synthetic Room Dataset. Learning-based methods are trained and tested on 10K noisy points. ALTO can reconstruct the (top scene) double-deck table and (bottom scene) details in the chair. Method IoU ↑ Chamfer-L1 ↓ NC↑ F-score↑
[]
ALTO achieves detailed qualitative results with higher fidelity reconstructio...
ALTO achieves detailed qualitative results compared to baselines, reconstructing fine details such as the double-deck table and chair details with higher fidelity
image D achieves detailed qualitative results compared to baselines, reconstructing fine details such as the double-deck table and chair details with higher fidelity
Object Form
Detail
0.81
paper_0011/fig_7_p6::row_3