dedup-isc-ft-v107-score
float64 0.32
1
| uid
stringlengths 32
32
| text
stringlengths 0
32.5k
| paper_id
stringlengths 1
14
| original_image_filename
stringlengths 5
222
|
|---|---|---|---|---|
0.532897 |
aaab4ffc43d44311bd3f38df9b1beeb7
|
Truncated flow field
|
2012.10572
|
F06.jpg
|
0.513939 |
2931297d907a40d4bb8180656700af1f
|
The case λ < 0.
|
2012.10572
|
F07.jpg
|
0.475634 |
cbc92c716dca42ae97dc9528565e4676
|
The case λ = 0
|
2012.10572
|
F08.jpg
|
0.528082 |
7bb08c4450274759a81046e4cb27e4ac
|
Two-phase fluid
|
2012.10572
|
F09.jpg
|
0.457471 |
852905b20a1b4d1eb9e7a57c420dc5af
|
The system running in the convex cell C without touching the boundary. x_0 is the initial point. x_e is the end point.
|
2012.10573
|
f1.jpg
|
0.536263 |
344e06e26d7a45feba427f5c40b44efa
|
Path Control Task: The system starts from a point x_in in the polygon (possibly on a facet l_1) and leaves from a point x_out on the exit facet l_2 without touching the boundary.
|
2012.10573
|
f2.jpg
|
0.38627 |
2a58863954ff4772bdcfe0003b67a947
|
d_i(x) is the distance from x to the i^th boundary in the polygon
|
2012.10573
|
figure1.jpg
|
0.437994 |
a1e6313c498d41c79be544ae3e251295
|
Images of A30 in F154W , F169M and F172M showing the FUV halo around F154 and F169M images and the absence of it around F172M image.
|
2012.10574
|
A30FuvF2F3F5dirs30BmK.jpg
|
0.412169 |
2a4415aa6854467eb1a339bc95690e33
|
UVIT/FUV F169M image of NGC 2818 is compared with [O iii] 5007 Å ground based, image (inset, Vazquez, 2012). While several features are in common, a major difference is the prominence of radial streamers (or emission filaments) in the F169M image, marked by green arrows. The bandpass of F169M filter covers mainly He ii λ1640 Å, and C iv λ 1550 Å emission lines. Presence of these streamers suggest that they might have been swept up by strong stellar wind from the central star.
|
2012.10574
|
Fuv169Ms5m1.19AOIIIn2818.jpg
|
0.452122 |
683d01887f554e948a8d9ab44f9dcd5e
|
IUE low resolution nebular spectra of NGC6302 are shown to illustrate the wavelength range of UVIT filters. FUV is plotted on top and NUV at the bottom. Relative effective areas of various UVIT filters used for PN studies and typical nebular emission lines they include are shown.
|
2012.10574
|
NGC6302IUESW1.jpg
|
0.525874 |
9d79d68e6a9745afa2c2b95e1c912909
|
Details of UVIT image of NGC 7293 in the N263M filter. There are some new structures that are not seen in optical images.
|
2012.10574
|
NGC7293NUV3neglogCAK.jpg
|
0.413752 |
33fe73a7057a44e19e970afa17d7919c
|
A surface and contour study of NGC 7293 in UVIT/FUV F169 and F172M bands. Top: Surface contours of the central source in F169 (left) and F172M (right) showing distored contours for F169M indicating the possible presence of He ii bright clouds or clumps around the CSPN. In F172M the contours more circular and suggests the presence of CSPN alone. Bottom: The figure shows the volume contours of the central source in F169M (He ii emission line) and F172M (continuum) filters. The CSPN is seen to be accompanied by three He ii clumps in F169M where as the profile in F172M shows CSPN alone.
|
2012.10574
|
NGC7293subco0KA.jpg
|
0.413052 |
92efc3b45cd74ab0b875fc8dc120f3a8
|
(Top) :NUV N219M image (left) is compared with UVIT image in F169M (F3) of NGC 2440.
|
2012.10574
|
NnuvB13fuvF3Cx5.18m.jpg
|
0.457901 |
9ef461ae3b4e4809b913d0d23b13aa6a
|
Details of UVIT image of NGC 7293 in the N263M filter. The fine optical knots detected in HST image (left) (O'Dell 2005) are seen UV image in F263M (right) obtained with UVIT. The same knots are shown in both images. Some of the cometary knots are also present in UV images.
|
2012.10574
|
fg2hodell2004PNUVB4.jpg
|
0.39663 |
64c17bc81b814cbaba646888007c7f65
|
Depiction of Commitment Transaction Stored at the LN Gateway with Fee Output Modified after IoT device sends 1 BTC payment to a destination. “Bridge" is Short for Bridge LN Node, “Gateway" is Short for the LN Gateway. Here “Output 4" is the Modified Output where the LN Gateway Charged a Fee of 0.1 BTC.
|
2012.10576
|
HTLC5.jpg
|
0.424556 |
94f1686d464f4a78a6892b350ee16852
|
Life cycle of an LN channel.
|
2012.10576
|
background.jpg
|
0.3759 |
26ddcbec036a43bb8f638bdd52515595
|
Protocol Messages for the IoT Device to close the channel.
|
2012.10576
|
closechannel5.jpg
|
0.488153 |
111dc0615fd543369f9bf313785f8362
|
Protocol Messages for Opening a Channel. Messages in red show the default messages in BOLT #2.
|
2012.10576
|
openchannel5.jpg
|
0.411531 |
2ec9031b0d3841f8aade9d8c3aa9a411
|
Protocol Messages for Sending a Payment. Messages in red show the default messages in BOLT #2.
|
2012.10576
|
sendpayment9.jpg
|
0.474664 |
a3f93f8094a04427bf0ea65dcab5f7f1
|
Illustration of our assumed system model.
|
2012.10576
|
system_model_3.jpg
|
0.449908 |
bf15d3801e4744aab0f180123a3bb02b
|
The fake images respectively from Faceforencis++ <cit.> (the 1st row) and the Celeb-DF <cit.> (the 2nd row). The ones in the 1st row exhibit the obvious artifacts, corresponding to the training data from ^L (in Def. <ref>) that we access. The 2nd row remove the artifacts by implementing better synthesis method (^H in Def. <ref>), as the test arena to validate the robustness of proposed methods.
|
2012.10580
|
fake_image3.jpg
|
0.472288 |
35872aa5aabe4b62919a7b578fdce9c6
|
Our InTeLe framework that learns from ^L and aims at learning the invariance with texture violation, i.e., p(y|z) to generalize on ^H. The InTeLe is composed of three modules: (i) auto-encoder with encoder Enc to extract the texture information and two-branch decoder (i.e., f_0, f_1) corresponding to generating process of pristine/fake images; (ii) the invariant classifier f_cls which exploits the information of the texture inconsistency; (iii) the semantic classification f_SC to encode the artifact-effect, by classifying f_0(Enc(x)) as 0 and f_1(Enc(x)) as 1; besides, it trains on (x,y) to regularize f_0,f_1 not to learn trivial solutions.
|
2012.10580
|
framework6.jpg
|
0.448051 |
26c22b4d0225472eaae4bbea937b50b8
|
Separation of artifacts by the two-branch decoder in our InTeLe. The top left, top right, bottom left and the bottom right quadrants correspond to pristine images in Celeb-DF, fake images in Celeb-DF, pristine images in benchmark and the fake images in benchmark. From the left to right in each row corresponds to original image x, reconstructed image by f_0 (i.e., f_0(Enc(x)))) and reconstructed image by f_1 (i.e., f_1(Enc(x))). As marked by the red circle, the f_1(Enc(x))) can expose obvious artifacts whether x is pristine/fake.
|
2012.10580
|
image_recon3.jpg
|
0.452179 |
3a2c25420a53411a86638eaf3e7dc6dc
|
The overall structure of the Adn.
|
2012.10820
|
Adn1.jpg
|
0.475391 |
ceee1a54f8c1443c9e9c4d2eac35a29f
|
The overall structure of our proposed AdnFM. The AdnFM includes two main components: an Adn component and a FM component. The Adn component learns the hierarchical attention weighted between different layers through an attention neural network. Its outputs of high-order features are integrated via a weighted pooling layer.
|
2012.10820
|
model.jpg
|
0.510151 |
f2673785d41b4e4faf091b234f581e71
|
Online A/B test results in Tencent game honor of kings.
|
2012.10820
|
wangzhe.jpg
|
0.541282 |
c848081e950d46d5b36bbfe69505a1d1
|
The evolution of the attention weights of dense layers during training.
|
2012.10820
|
weight.jpg
|
0.493713 |
0ad25a62efb344a3af3358e945fe92e5
|
The generated region proposals with different proposal generation methods on COCO. (a) Selective Search (SS); (b) Faster-RCNN pre-trained on COCO; (c) Our phrase-guided proposal generation network.The red boxes represent the proposals, and the green boxes represent the ground truth.
|
2012.10890
|
pic1.jpg
|
0.416582 |
e8262a15ea2f450f9f716d989a2081bd
|
Overall structure of our model. The part of stage 1 is our phrase-guided proposal generation network model, and the part of stage 2 is proposal representation and ranking module. These two parts are combined to make up a complete REC model.
|
2012.10890
|
pic2.jpg
|
0.373696 |
d7d1d282f044446a8a81b9f65d5e2756
|
The example in Referit dataset.The ground-truth(green), the final top-ranked predicted proposal(red), and the other proposals (blue) are visualized respectively. For better viewing, we removed some highly overlapping bounding boxes.We also show 3 examples of incorrect predictions(IoU ≤ 0.5).Best viewed in color.
|
2012.10890
|
pic3.jpg
|
0.352952 |
d78d091a600249be907276a4afe7128e
|
Amplitudes of perturbations evaluated at y=0 in the evolution of initial data (<ref>) for different values of the wormhole radius: a=0.25 (no gap eigenvalue) and a=1 (ω=1.0682), a=2 (ω=0.6345), a=3 (ω=0.4455), corresponding to N=1,2,3 in (<ref>), respectively. The red dashed lines depict analytic predictions with empirically fitted coefficients.
|
2012.11141
|
decay_logh.jpg
|
0.449669 |
afee6224fe604360882f19a6d28ec687
|
The frequency of the internal mode of the 1-kink as a function of 1/a.
|
2012.11141
|
omega.jpg
|
0.503603 |
5231ca2c084847169da7ab057e38e9a5
|
The snapshots of h(s,y) for the initial data (<ref>). The solution converges to the 1-kink h_1(y) (solid black line) in an oscillatory manner.
|
2012.11141
|
snapshot_a1.jpg
|
0.487998 |
f4de5853edd148f79a230ad4dd6b97f8
|
Schematic of ONN on-chip learning framework with stochastic zeroth-order mixed-training.
|
2012.11148
|
Framework.jpg
|
0.445798 |
ae6b2241b9e243069eaf4b637579c3ab
|
MZI triangular array U(n) and the MZI structure.
|
2012.11148
|
MZITriangularArray.jpg
|
0.478298 |
25ae0ecf63c842abb8e1220530c1f785
|
Mixed-training flow in the practical ONN deployment.
|
2012.11148
|
MixedTrainMerge.jpg
|
0.407791 |
aea8b7c9663644ea916dbffbe0d6f90c
|
2012.11149
|
BP_model_smooth.jpg
|
|
0.408909 |
b4a16713d5d34cb4b94e13eb164b5469
|
2012.11149
|
BP_model_true.jpg
|
|
0.396236 |
1a74de376bc74a4bb9b3e54a717e33fa
|
2012.11149
|
FWI_BP.jpg
|
|
0.393793 |
acdff0555baf47c5b440e8dad1d33842
|
2012.11149
|
FWI_BP_1.jpg
|
|
0.420821 |
483f0c2691614f8e97b5f9ea58df8e1e
|
2012.11149
|
FWI_BP_24.jpg
|
|
0.400148 |
e9a873a0c0dc42919c77342c8ec2effe
|
2012.11149
|
FWI_BP_5.jpg
|
|
0.442087 |
523a1f8b648f43dc809be2a774e7b31e
|
2012.11149
|
FWI_BP_TV_2.jpg
|
|
0.431271 |
8dac38fe4d3049eca406c4c6019b13d1
|
2012.11149
|
FWI_BP_TV_3.jpg
|
|
0.393334 |
ecfcc6e246a64cdfaee498a3f94b6978
|
2012.11149
|
FWI_BP_TV_3_1.jpg
|
|
0.401972 |
f3f00e05b8cc441288651e77700dda04
|
2012.11149
|
FWI_BP_TV_3_33.jpg
|
|
0.384239 |
9b36d762f6c14149ae7609d34655f3f2
|
2012.11149
|
FWI_BP_TV_3_6.jpg
|
|
0.417495 |
052f6a8cb74c4bfe880e91dba1747b89
|
2012.11149
|
FWI_BP_TV_4.jpg
|
|
0.411095 |
c0ba47b3425d464382f1b19ef2724153
|
2012.11149
|
FWI_BP_adam.jpg
|
|
0.523199 |
0cf797fb5abf42cf84ff704ef6b268b6
|
2012.11149
|
FWI_marmousi.jpg
|
|
0.410788 |
d7c9aba7e56e4481b14b6256302758c0
|
2012.11149
|
FWI_marmousi_1.jpg
|
|
0.429152 |
5ec6ef0d01b04964ab5074467f0c5be9
|
2012.11149
|
FWI_marmousi_2.jpg
|
|
0.436251 |
4178258295db40338d05074db9d89120
|
2012.11149
|
FWI_marmousi_3.jpg
|
|
0.512748 |
d21aea90fb294191a9afd1f54d2af7a2
|
2012.11149
|
FWI_marmousi_adam.jpg
|
|
0.425745 |
cf31ed75c9834487a5e3d1537cb2cc76
|
2012.11149
|
FWI_marmousi_noise05.jpg
|
|
0.41843 |
9c3ed8dae23a41fd8d8e71df4d5c55f7
|
2012.11149
|
FWI_marmousi_noise10.jpg
|
|
0.446358 |
ff298a07916c450ea744c3935c34edd8
|
(a) Workflow of conventional FWI based on the adjoint-state method. (b) Diagram of NNFWI, which combines neural networks and PDEs of seismic waves. The input velocity to the PDEs is a direct summation of the initial velocity model and the velocity model generated by the neural network. In this way, NNFWI follows the same data format as conventional FWI and can be directly applied to conventional FWI applications. The gradients of both the neural network and PDE are calculated using automatic differentiation in NNFWI using the ADSeismic package <cit.>.
|
2012.11149
|
NNFWI.jpg
|
0.381325 |
15c4d51c852d4e9d92519c7e26b62972
|
2012.11149
|
NNFWI_BP.jpg
|
|
0.463233 |
714873e19d8f440ea2a603392e23a2df
|
2012.11149
|
NNFWI_BP_23.jpg
|
|
0.435191 |
09b68e254345495da2e5216d4da4e5fc
|
2012.11149
|
NNFWI_BP_57.jpg
|
|
0.44207 |
a4a9fba8f6ab4ef2bf503269f9f7f90f
|
2012.11149
|
NNFWI_BP_9.jpg
|
|
0.40675 |
be1de29fb3e5482aa720ec64bf194553
|
2012.11149
|
NNFWI_BP_BFGS.jpg
|
|
0.394729 |
7bc4a01d40de409dbf40ad9332f36e9e
|
2012.11149
|
NNFWI_BP_UQ.jpg
|
|
0.389404 |
20833f88125c4aac94dbde58f46595fb
|
2012.11149
|
NNFWI_BP_UQ_3.jpg
|
|
0.411359 |
00b248ebf6f3431fad8e1faf7db26c1d
|
2012.11149
|
NNFWI_BP_UQ_3_err.jpg
|
|
0.4387 |
42cda039d73b42439b3133ef1cb3f5bd
|
2012.11149
|
NNFWI_BP_UQ_3_profile.jpg
|
|
0.3964 |
7f86e96c2d374c39ba88d6ae3770a0a4
|
2012.11149
|
NNFWI_BP_UQ_3_std.jpg
|
|
0.432664 |
fc4b140ec1c2479e942d5cd2f873f311
|
2012.11149
|
NNFWI_BP_UQ_dp2.jpg
|
|
0.469342 |
3241de2394d44d748abc6c317577b046
|
2012.11149
|
NNFWI_BP_UQ_dp2_err.jpg
|
|
0.447159 |
df32823426ee4a6e8ca4bf1f6ded5b71
|
2012.11149
|
NNFWI_BP_UQ_dp2_profile.jpg
|
|
0.411055 |
c29cdc8d3d754bdca4059a0f3adbeb37
|
2012.11149
|
NNFWI_BP_UQ_dp2_std.jpg
|
|
0.442497 |
96f3ad955a2842b3a84ffa48a2cfd75e
|
2012.11149
|
NNFWI_BP_UQ_err.jpg
|
|
0.449874 |
6889bf2122d64c118c751a252c6c5141
|
2012.11149
|
NNFWI_BP_UQ_profile.jpg
|
|
0.461448 |
927cffcdf39e4a489b3f52b06d0beac4
|
2012.11149
|
NNFWI_BP_UQ_std.jpg
|
|
0.478863 |
bc109a0aa8f6408ea3986a950e6af053
|
2012.11149
|
NNFWI_marmousi.jpg
|
|
0.436788 |
feabfa90fc7b41c39103c2986de96bd3
|
2012.11149
|
NNFWI_marmousi_1.jpg
|
|
0.419528 |
3f8320aff86449c4992d42384a7744c6
|
2012.11149
|
NNFWI_marmousi_27.jpg
|
|
0.415792 |
a1249b5a296f46788813ac22d1ac9878
|
2012.11149
|
NNFWI_marmousi_89.jpg
|
|
0.452185 |
6435fa5bd2bc4ff8bbbaa1c94482f3bd
|
2012.11149
|
NNFWI_marmousi_BFGS.jpg
|
|
0.460819 |
c77aada4c6c544b3a65d2adc2cee3e63
|
2012.11149
|
NNFWI_marmousi_UQ.jpg
|
|
0.50807 |
95043d650a3940b0b7511a5822dc77dc
|
2012.11149
|
NNFWI_marmousi_UQ_3.jpg
|
|
0.442086 |
217747fee8db4d999aa4de8a8b027e80
|
2012.11149
|
NNFWI_marmousi_UQ_3_err.jpg
|
|
0.396458 |
9f8b64664ff74775b4e08bade5d411d4
|
2012.11149
|
NNFWI_marmousi_UQ_3_profile.jpg
|
|
0.472097 |
6ff52b7869bb425d8df84a80deeb4079
|
2012.11149
|
NNFWI_marmousi_UQ_3_std.jpg
|
|
0.484858 |
cb9c019724264dd19d18b3975a7ad413
|
2012.11149
|
NNFWI_marmousi_UQ_dp2.jpg
|
|
0.444248 |
3bd19670da29472095a3ff2a71be3615
|
2012.11149
|
NNFWI_marmousi_UQ_dp2_err.jpg
|
|
0.411059 |
f2ee304633974856bfff2e56663e8f4c
|
2012.11149
|
NNFWI_marmousi_UQ_dp2_profile.jpg
|
|
0.423723 |
11db7834208143f494fe6a093c60d145
|
2012.11149
|
NNFWI_marmousi_UQ_dp2_std.jpg
|
|
0.448458 |
c6335b128e4a4165bbf6016b1d43b1f7
|
2012.11149
|
NNFWI_marmousi_UQ_err.jpg
|
|
0.406141 |
f41ad7cee9a7422b9f2907d74ac1f4be
|
2012.11149
|
NNFWI_marmousi_UQ_profile.jpg
|
|
0.437031 |
3aa979f48c8d42ada9ac340845e19840
|
2012.11149
|
NNFWI_marmousi_UQ_std.jpg
|
|
0.496429 |
d36afc425e384b689cf00964cf0f3683
|
2012.11149
|
NNFWI_marmousi_noise05.jpg
|
|
0.415272 |
5c093d2c522c4eadac871ffc21c8db58
|
2012.11149
|
NNFWI_marmousi_noise10.jpg
|
|
0.500823 |
a5cac796339c4b69919be209bd696b6b
|
Inversion results of an 1D linear velocity profile. The upper panel (a-c) shows results of conventional FWI. The middle panel (d-f) shows results of FWI with Tikhonov regularization. The lower panel (g-i) shows results of NNFWI. The true and estimated velocity profiles are plotted in the left panel (a, d, g). We run 10,000 interactions for the final estimation. The estimated profiles at three selected iterations are also plotted to show the convergence. Note that we only estimate the velocity at x≥0.1 km to avoid extreme gradient values at the source. Because we only placed two receivers at x=0 and x=1.0, the updates of the first few iterations of FWI mainly occurred in the area between the two receivers, causing a box-like perturbation in the first iteration of FWI in panel (a). NNFWI, however, imposes a spatial regularization to update the whole region, as shown in panel (g). The fitted waveforms are plotted in the middle panel (b, e, h). The whole wavefields are plotted in the right panel (c, f, i).
|
2012.11149
|
compare_slop_legend.jpg
|
0.414197 |
01d4b87b7af34814b0a49de8afc30fcb
|
Inversion results of an 1D step-change velocity profile. The panels are plotted in the same way as <Ref>. Note that the initial step-change at x=0.1 km occurs because we only estimate the velocity at x≥0.1 km and fix the true velocity value at x<0.1 km to avoid source effects. The waveforms at two receivers are well fit for all three methods, but we observe significant differences in the wavefields (c, f, i) between the two receivers due to the incorrect velocity models.
|
2012.11149
|
compare_step_legend.jpg
|
0.410823 |
360f7a0b73984cd78a6823444a682bdd
|
2012.11149
|
loss_BP.jpg
|
|
0.44763 |
536162f8a7104e1cb1f3b1015801641d
|
2012.11149
|
loss_marmousi.jpg
|
|
0.447351 |
d3db663527ea4f5eb0fda0abe2649303
|
2012.11149
|
loss_marmousi_layers.jpg
|
|
0.43107 |
91668e393a544800a4f352ca3a681807
|
2012.11149
|
loss_marmousi_noise05.jpg
|
|
0.384953 |
ab2a09bc72c74ec68e04b8ba5f44ebb4
|
2012.11149
|
marmousi_model_smooth.jpg
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.