@@ -124,6 +124,7 @@ <h2 class="title is-3">Abstract</h2>
124124 changes in reconstructed images, revealing that orthogonal rotations — rather than spatial transformations
125125 — control color encoding. Our approach can be applied to any vision encoder, shedding light on the inner
126126 structure of its feature space.
127+ </ p >
127128 </ div >
128129 </ div >
129130 </ div >
@@ -139,26 +140,33 @@ <h2 class="title is-3">Abstract</h2>
139140 <!-- Нужно четко подчеркнуть какие различия между модельками -->
140141 <!-- И как это влияет на реконструкцию -->
141142
142- < div class ="columns is-centered has-text-centered ">
143- < div class ="column is-four-fifths ">
144- < h2 class ="title is-3 " style ="white-space: nowrap; "> Reconstruct images from feature space</ h2 >
145- < div class ="content has-text-justified ">
146- < div style ="text-align: center; ">
147- < img src ="./static/images/features_reconstruction.drawio.png " alt ="features_reconstruction " width ="900 ">
148- < p class ="caption " style ="width: 100%; text-align: center; ">
149- < b > Figure 1. Image reconstructor training.</ b > For pretrained model we train a reconstructor model that
150- restores the image from the feature space.
151- </ p >
152- < img src ="./static/images/reconstruction_metrics.jpg " alt ="reconstruction_metrics " width ="900 ">
153- < p class ="caption " style ="width: 100%; text-align: center; ">
154- < b > Figure 2. Reconstruction Metrics.</ b > We show the results of the reconstruction for SigLip and SigLip2
155- for different image resultions.
156- </ p >
143+ < section class ="section ">
144+ < div class ="container is-max-desktop ">
145+ < div class ="columns is-centered ">
146+ < div class ="column is-four-fifths ">
147+ < h2 class ="title is-3 has-text-centered "> Reconstruct images from feature space</ h2 >
148+ < div class ="content ">
149+ < div class ="has-text-centered ">
150+ < figure class ="image is-16by9 mb-6 ">
151+ < img src ="./static/images/features_reconstruction.drawio.png " alt ="features_reconstruction " class ="is-rounded ">
152+ </ figure >
153+ < p class ="caption has-text-centered mb-6 ">
154+ < b > Figure 1. Image reconstructor training.</ b > For pretrained model we train a reconstructor model that
155+ restores the image from the feature space.
156+ </ p >
157+ < figure class ="image is-16by9 mb-6 ">
158+ < img src ="./static/images/reconstruction_metrics.jpg " alt ="reconstruction_metrics " class ="is-rounded ">
159+ </ figure >
160+ < p class ="caption has-text-centered mb-6 ">
161+ < b > Figure 2. Reconstruction Metrics.</ b > We show the results of the reconstruction for SigLip and SigLip2
162+ for different image resultions.
163+ </ p >
164+ </ div >
165+ </ div >
157166 </ div >
158167 </ div >
159- < br > < br >
160168 </ div >
161- </ div >
169+ </ section >
162170
163171 <!-- (2) Feature-space transformations -->
164172 <!-- Текстовое объяснение -->
@@ -168,90 +176,115 @@ <h2 class="title is-3" style="white-space: nowrap;">Reconstruct images from feat
168176 <!-- Примеры Спектра такой м-цы, показать, что только небольшое кол-во каналов меняется -->
169177 <!-- -->
170178
171- < div class ="columns is-centered has-text-centered ">
172- < div class ="column is-four-fifths ">
173- < h2 class ="title is-3 " style ="white-space: nowrap; "> Feature-space transformations. Q matrix Calculation and Application.</ h2 >
174- < div class ="content has-text-justified ">
175- < div class ="columns is-centered ">
176- < div class ="column is-half ">
177- < img src ="./static/images/features_reconstruction_manipulation_train_Q.drawio.png " alt ="features_reconstruction_manipulation_train_Q " width ="100% ">
178- < p class ="caption " style ="width: 70%; text-align: center; ">
179- < b > Figure 3. Feature-space transformations. Q matrix Calculation.</ b > We then calculate Q matrix for feature-space manupulation.
180- </ p >
181- </ div >
182- < div class ="column is-half ">
183- < img src ="./static/images/features_reconstruction_manipulation_eval_Q.drawio.png " alt ="features_reconstruction_manipulation_eval_Q " width ="100% ">
184- < p class ="caption " style ="width: 70%; text-align: center; ">
185- < b > Figure 4. Feature-space transformations. Q matrix Application.</ b > After Q matrix is calculated, we apply it to the feature space. For each patch embedding.
186- </ p >
179+ < section class ="section ">
180+ < div class ="container is-max-desktop ">
181+ < div class ="columns is-centered ">
182+ < div class ="column is-four-fifths ">
183+ < h2 class ="title is-3 has-text-centered "> Feature-space transformations. Q matrix Calculation and Application.</ h2 >
184+ < div class ="content ">
185+ < div class ="columns is-centered ">
186+ < div class ="column is-half ">
187+ < figure class ="image is-16by9 mb-6 ">
188+ < img src ="./static/images/features_reconstruction_manipulation_train_Q.drawio.png " alt ="features_reconstruction_manipulation_train_Q " class ="is-rounded ">
189+ </ figure >
190+ < p class ="caption has-text-centered mb-6 ">
191+ < b > Figure 3. Feature-space transformations. Q matrix Calculation.</ b > We then calculate Q matrix for feature-space manupulation.
192+ </ p >
193+ </ div >
194+ < div class ="column is-half ">
195+ < figure class ="image is-16by9 mb-6 ">
196+ < img src ="./static/images/features_reconstruction_manipulation_eval_Q.drawio.png " alt ="features_reconstruction_manipulation_eval_Q " class ="is-rounded ">
197+ </ figure >
198+ < p class ="caption has-text-centered mb-6 ">
199+ < b > Figure 4. Feature-space transformations. Q matrix Application.</ b > After Q matrix is calculated, we apply it to the feature space. For each patch embedding.
200+ </ p >
201+ </ div >
202+ </ div >
187203 </ div >
188204 </ div >
189205 </ div >
190- < br > < br >
191206 </ div >
192- </ div >
207+ </ section >
193208
194- < div class ="columns is-centered has-text-centered ">
195- < div class ="column is-four-fifths ">
196- < h2 class ="title is-3 " style ="white-space: nowrap; "> Feature-space transformations. Color Swap Examples.</ h2 >
197- < div class ="content has-text-justified ">
198- < div class ="columns is-centered ">
199- < div class ="column is-half ">
200- < img src ="./static/images/rb_swap.png " alt ="rb_swap " width ="100% ">
201- < p class ="caption " style ="width: 100%; text-align: center; ">
202- < b > Figure 5. Red-blue channel swap samples.</ b >
203- </ p >
204- </ div >
205- < div class ="column is-half ">
206- < img src ="./static/images/color_swap_all_eigen_values.png " alt ="color_swap_all_eigen_values " width ="100% ">
207- < p class ="caption " style ="width: 100%; text-align: center; ">
208- < b > Figure 6. Eigenvalues for red-blue channel swap matrix.</ b > Majority of eigenvalues are close to 1, which means that the transformation is close to an identity matrix. While the other cluster of eigenvalues are close to -1, which means that for these channels direction is changed to the opposite.
209- </ p >
209+ < section class ="section ">
210+ < div class ="container is-max-desktop ">
211+ < div class ="columns is-centered ">
212+ < div class ="column is-four-fifths ">
213+ < h2 class ="title is-3 has-text-centered "> Feature-space transformations. Color Swap Examples.</ h2 >
214+ < div class ="content ">
215+ < div class ="columns is-centered ">
216+ < div class ="column is-half ">
217+ < figure class ="image is-16by9 mb-6 ">
218+ < img src ="./static/images/rb_swap.png " alt ="rb_swap " class ="is-rounded ">
219+ </ figure >
220+ < p class ="caption has-text-centered mb-6 ">
221+ < b > Figure 5. Red-blue channel swap samples.</ b >
222+ </ p >
223+ </ div >
224+ < div class ="column is-half ">
225+ < figure class ="image is-16by9 mb-6 ">
226+ < img src ="./static/images/color_swap_all_eigen_values.png " alt ="color_swap_all_eigen_values " class ="is-rounded ">
227+ </ figure >
228+ < p class ="caption has-text-centered mb-6 ">
229+ < b > Figure 6. Eigenvalues for red-blue channel swap matrix.</ b > Majority of eigenvalues are close to 1, which means that the transformation is close to an identity matrix. While the other cluster of eigenvalues are close to -1, which means that for these channels direction is changed to the opposite.
230+ </ p >
231+ </ div >
232+ </ div >
210233 </ div >
211234 </ div >
212235 </ div >
213- < br > < br >
214236 </ div >
215- </ div >
237+ </ section >
216238
217- < div class ="columns is-centered has-text-centered ">
218- < div class ="column is-four-fifths ">
219- < h2 class ="title is-3 " style ="white-space: nowrap; "> Feature-space transformations. Blue Channel Suppression.</ h2 >
220- < div class ="content has-text-justified ">
221- < div class ="columns is-centered ">
222- < div class ="column is-half ">
223- < img src ="./static/images/b_suppression_all_transformations.png " alt ="b_suppression_all_transformations " width ="100% ">
224- < p class ="caption " style ="width: 100%; text-align: center; ">
225- < b > Figure 7. Blue channel suppression samples.</ b >
226- </ p >
227- </ div >
228- < div class ="column is-half ">
229- < img src ="./static/images/b_suppression_all_eigen_values.png " alt ="b_suppression_all_eigen_values " width ="100% ">
230- < p class ="caption " style ="width: 100%; text-align: center; ">
231- < b > Figure 8. Eigenvalues for blue channel suppression matrix.</ b >
232- </ p >
239+ < section class ="section ">
240+ < div class ="container is-max-desktop ">
241+ < div class ="columns is-centered ">
242+ < div class ="column is-four-fifths ">
243+ < h2 class ="title is-3 has-text-centered "> Feature-space transformations. Blue Channel Suppression.</ h2 >
244+ < div class ="content ">
245+ < div class ="columns is-centered ">
246+ < div class ="column is-half ">
247+ < figure class ="image is-16by9 mb-6 ">
248+ < img src ="./static/images/b_suppression_all_transformations.png " alt ="b_suppression_all_transformations " class ="is-rounded ">
249+ </ figure >
250+ < p class ="caption has-text-centered mb-6 ">
251+ < b > Figure 7. Blue channel suppression samples.</ b >
252+ </ p >
253+ </ div >
254+ < div class ="column is-half ">
255+ < figure class ="image is-16by9 mb-6 ">
256+ < img src ="./static/images/b_suppression_all_eigen_values.png " alt ="b_suppression_all_eigen_values " class ="is-rounded ">
257+ </ figure >
258+ < p class ="caption has-text-centered mb-6 ">
259+ < b > Figure 8. Eigenvalues for blue channel suppression matrix.</ b >
260+ </ p >
261+ </ div >
262+ </ div >
233263 </ div >
234264 </ div >
235265 </ div >
236- < br > < br >
237266 </ div >
238- </ div >
239-
240- < div class ="columns is-centered has-text-centered ">
241- < div class ="column is-four-fifths ">
242- < h2 class ="title is-3 " style ="white-space: nowrap; "> Feature-space transformations. Colorization.</ h2 >
243- < div class ="content has-text-justified ">
244- < div style ="text-align: center; ">
245- < img src ="./static/images/colorized_examples.png " alt ="colorization_all_transformations " width ="900 ">
246- < p class ="caption " style ="width: 100%; text-align: center; ">
247- < b > Figure 9. Colorization samples.</ b >
248- </ p >
267+ </ section >
268+
269+ < section class ="section ">
270+ < div class ="container is-max-desktop ">
271+ < div class ="columns is-centered ">
272+ < div class ="column is-four-fifths ">
273+ < h2 class ="title is-3 has-text-centered "> Feature-space transformations. Colorization.</ h2 >
274+ < div class ="content ">
275+ < div class ="has-text-centered ">
276+ < figure class ="image is-16by9 mb-6 ">
277+ < img src ="./static/images/colorized_examples.png " alt ="colorization_all_transformations " class ="is-rounded ">
278+ </ figure >
279+ < p class ="caption has-text-centered mb-6 ">
280+ < b > Figure 9. Colorization samples.</ b >
281+ </ p >
282+ </ div >
283+ </ div >
249284 </ div >
250285 </ div >
251286 </ div >
252- < br > < br >
253- </ div >
254- </ div >
287+ </ section >
255288
256289
257290
0 commit comments