image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 4:Example images from the Dynamic Traffic Speeds (DTS) dataset and corresponding labels. The right four labels depict available historical traffic speeds at different times, where green (red) corresponds to faster (slower) speeds. | 2403.05521v2 | 4 |
|
Figure 5:Qualitative results for traffic speed estimation (Monday, 8am). (top) Ground-truth traffic speeds are provided as aggregates for individual road segments. For visualization, we replicate the ground-truth speed across the entire road segment. (middle) Predictions from our approach, after applying region aggregation. (bottom) Our results without region aggregation capture nuances of traffic flow, such as slowing down around curves. Green (red) corresponds to fast (slow) traffic speeds. | 2403.05521v2 | 7 |
|
Figure 6:Visualizing traffic speed predictions for individual road segments versus time. (left) A road segment shown in red. (right) Predictions from our approach (red) versus historical traffic data (green). | 2403.05521v2 | 8 |
|
Figure 7:The output of our approach on unseen locations visualized as a local motion model. We sample from locations predicted as road and visualize corresponding orientation estimates as vectors, colored by the predicted speed. Green (red) corresponds to fast (slow) traffic speeds. | 2403.05521v2 | 9 |
|
Figure S1:Example traffic speed data from DTS++ for Cincinnati (road segment speeds averaged over time). Green (red) corresponds to fast (slow) traffic speeds. | 2403.05521v2 | 10 |
|
Figure S2:(left) Visualizing the average road segment speed over time in Cincinnati for different days. (right) Visualizing the average road segment speed against time in Cincinnati, where each road is categorized by its OpenStreetMap road type classification. | 2403.05521v2 | 11 |
|
Figure S3:Visualizing the spatiotemporal prior, learned using the GTPE module, for several species in the iNaturalist 2018 dataset. Darker colors correspond to higher likelihood of the species appearing at that location. | 2403.05521v2 | 12 |
|
Figure S5:Our method outputs prior distributions over traffic speeds that implicitly capture uncertainty. (right) Output from our approach visualized as a probability density function corresponds to the (left) road segment depicted in the image. The greenx𝑥xitalic_xrepresents the ground-truth traffic speed at the given time. | 2403.05521v2 | 14 |
|
Figure S6:Qualitative results for the auxiliary tasks of road segmentation and orientation estimation. Orientation is represented byK=16𝐾16K=16italic_K = 16bins using the HSV color map. | 2403.05521v2 | 15 |
|
Figure S7:The output of our approach can be used for various applications, such as generating isochrone maps reflecting travel time. This map reflects travel times on a Monday at 8am. Starting from the center, each band reflects areas reachable in five minute increments (dark red shortest, dark blue furthest). | 2403.05521v2 | 16 |
|
Figure 1.Schematic diagram of the proposed system. The first two steps are language modelling tasks, and the last step is authorship attribution. The embedding layer and the AWD-LSTM base remain the same in all steps, only the final dense layer changes. The trained weights of the fixed parts are passed on from pre-training to fine-tuning to the classification step, updated in each step. | 2403.05519v1 | 1 |
|
Figure 3.Example of three kinds of tokenization of a sample text. Punctuation added for demonstration purposes. | 2403.05519v1 | 5 |
|
Figure 7.Application of Discriminative Learning Rates(Lynn-Evans,2019). The initial layers (red) have a lower learning rate to retain the pre-trained weights, while later layers (green) have a higher learning rate to tune faster to the task at hand. | 2403.05519v1 | 14 |
|
Figure 8.Resetting Learning Rate after each Epoch and reducing gradually by cosine annealing(Lynn-Evans,2019) | 2403.05519v1 | 15 |
|
Figure 10.The accuracy versus epoch plot shows the models’ performance comparison with different tokenizations. Here the character tokenized model cannot reach the level of subword and word tokenized models in both the tested datasets (BAAD6 and BAAD16). Notice the vertical axis does not start from zero for clarity. | 2403.05519v1 | 17 |
|
Figure 12.Performance of different models with increasing number of authors. The models vary by tokenization and the dataset it was pre-trained on, depicted in the legend in the form ’tokenization_pretraining-dataset’. The vertical axis shows the classification accuracy, and the horizontal axis indicates the number of authors taken from the BAAD16 dataset. | 2403.05519v1 | 22 |
|
Figure 1:Performing bias-augmented consistency training (BCT) withSuggested Answerreduces biased reasoning on held-out tasks and a wide range of held-out biases. BCT improves over the self-training control on all biases except forPositional Bias.
For clarity, we paraphrase the prompts andboldthe bias augmentations.
Final responses show the biased answer to the question.Bias %measures how often models answer in line with particular incorrect answers that we bias them towards.Are You SureandPositional Biasare measured differently (§3.4).
Theunbiased baselinemeasures how often the original model (i.e. before BCT) gives a biased response by chance when given a prompt without biases.
The difference in theBias %for a model and the unbiased baseline is a measure of unfaithful biased reasoning because models generally do not mention the influence of biases (§3.2). | 2403.05518v1 | 1 |
|
Figure 2:A depiction of bias-augmented consistency training. We generate unbiased CoT reasoning by querying a model with a standard prompt without biasing features. We add bias augmentations to create biased prompts.
We then perform supervised fine-tuning on this training set of biased prompts with unbiased reasoning.
The purple dashed arrow above denotes the target behavior. Responses are from GPT-3.5, paraphrased for brevity. | 2403.05518v1 | 2 |
|
Figure 6:Biased reasoning rates as a function of the proportion of the data mixture that is BCT data, as opposed to intruction-tuning data. Increasing the proportion of BCT data does not always further decrease biased reasoning. We hypothesize that this could be due to overfitting since we only train on the Suggested Answer bias. | 2403.05518v1 | 6 |
|
Figure 1:Conceptual visualization of (un)familiar finetuning examples (left), and example of model predictions mimicking unfamiliar finetuning examples (middle and right). When finetuning on distribution 1, which contains details the model may not know, the model outputs detailed responses at test-time with inaccuracies (red). When finetuning on distribution 2, which omits unfamiliar details, the model produces shorter responses with fewer inaccuracies. | 2403.05612v2 | 1 |
|
Figure 2:Prediction behavior of models finetuned with SFT on MMLU (top 2 rows) and TriviaQA (bottom row). For MMLU plots, only test inputs with a specific ground truth label (A-D) are evaluated within each column. Solid line represents the average predicted likelihood, and error bars represent standard deviation within the test set. For TriviaQA plots, each bar denotes the ratio of model outputs within each category. For all plots in this figure, as inputs become more unfamiliar, model predictions default towards the distribution of target responses in the model’s unfamiliar finetuning examples. | 2403.05612v2 | 2 |
|
Figure 3:Prediction behavior of models finetuned with RL on MMLU (left) and TriviaQA (right). As inputs become more unfamiliar, the models finetuned with the first reward function produced random guesses while models finetuned with the section reward function produced abstain answers. | 2403.05612v2 | 3 |
|
Figure 4:Prediction behavior of reward models finetuned on MMLU (left 2) and TriviaQA (right 2). Green line represents model predictions for test examples that are correct (reward 1), and red line represents predictions for incorrect examples (reward 0). As inputs become more unfamiliar, the reward models produce different kinds of hallucinations depending on their finetuning distribution. | 2403.05612v2 | 4 |
|
Figure 5:Average reward predicted by a standard reward model and a conservative reward model as inputs become more unfamiliar, as well as the average ground truth reward. The standard reward model tends to overestimate rewards as input become more unfamiliar, whereas the conservative reward model does not. | 2403.05612v2 | 5 |
|
Figure 5:Average reward predicted by a standard reward model and a conservative reward model as inputs become more unfamiliar, as well as the average ground truth reward. The standard reward model tends to overestimate rewards as input become more unfamiliar, whereas the conservative reward model does not. | 2403.05612v2 | 6 |
|
Figure 7:Average number of true and false facts generated by models finetuned with standard SFT, RL with a standard reward model, and RL with a conservative reward model, as inputs become more unfamiliar. The responses generated by model finetuned with s conservative reward model consisted of fewer false facts and and equal number or more truth facts. | 2403.05612v2 | 7 |
|
Figure 8:Examples of generated responses from models finetuned with standard SFT and RL with a conservative reward model. False information is highlighted in red. | 2403.05612v2 | 8 |
|
Figure 2:ComparingMIbounds with trueMIin the Gaussian setting.Each method is trained for200200200200with multiplicitiesM∈{2,4,8,10}𝑀24810M\in\{2,4,8,10\}italic_M ∈ { 2 , 4 , 8 , 10 }.
Left to right:
1) True One-vs-RestMI(Equation26);
2)MIGaps decrease asM𝑀Mitalic_Mgrows for all methods except Multi-Crop due to thelog(K)𝐾\log(K)roman_log ( italic_K )factor;
3) RelativeMI= True MI / Lower Bound MI;
and 4) losses for each objective.
Bands indicate the mean and standard deviation across16161616runs. Points indicate final model performance of corresponding hyperparameters. | 2403.05490v1 | 4 |
|
Figure 4:Content-Style causal graphA more general poly-view sample generation with task-related generative factor𝐜𝐜\mathbf{c}bold_c, calledcontent. For eachα∈[M]𝛼delimited-[]𝑀\alpha\in[M]italic_α ∈ [ italic_M ],𝐬αsubscript𝐬𝛼\mathbf{s}_{\alpha}bold_s start_POSTSUBSCRIPT italic_α end_POSTSUBSCRIPTshows the view-dependent and task non-related factors, calledstyles. The views are shown as before by𝐱αsubscript𝐱𝛼\mathbf{x}_{\alpha}bold_x start_POSTSUBSCRIPT italic_α end_POSTSUBSCRIPT. In the most general case, content and styles are not independent, while in some settings they might be independent. In the independent scenario, the arrows from𝐜𝐜\mathbf{c}bold_cto𝐬αsubscript𝐬𝛼\mathbf{s}_{\alpha}bold_s start_POSTSUBSCRIPT italic_α end_POSTSUBSCRIPTcan be ignored. | 2403.05490v1 | 8 |
|
Figure 1:JointMotion.(a)Connecting motion and environments: Our scene-level objective learns joint scene representations via non-contrastive similarity learning of motion sequencesM𝑀Mitalic_Mand environment contextE𝐸Eitalic_E.(b)Masked polyline modeling: Our instance-level objective refines learned representations via masked autoencoding of multimodal polyline embeddings (i.e., motion, lane, and traffic light data). | 2403.05489v2 | 1 |
|
Figure 2:Adaptive decoding for masked polyline modeling with late and early fusion encoders.(a)Latefusion with modality-specific encoders for agents (EncoderAA{}^{\text{A}}start_FLOATSUPERSCRIPT A end_FLOATSUPERSCRIPT), lanes (EncoderLL{}^{\text{L}}start_FLOATSUPERSCRIPT L end_FLOATSUPERSCRIPT), and traffic lights (EncoderTLTL{}^{\text{TL}}start_FLOATSUPERSCRIPT TL end_FLOATSUPERSCRIPT).(b)Early fusion with a shared encoder for all modalities. Compressed features are decoded using learned query tokens. | 2403.05489v2 | 2 |
|
Figure 1:Architecture of our system. The core piece of our code, the Manager, is a Matplotlib interface able to run Doom on top of the Python binding, itself an interface for the original Doom code in C.
The Manager communicates states and actions to the binding; sends screenshots to Vision, and retains and parses the history of previous moves by Agent, Planner and Experts (if applicable).
The three boxes above describe our prompt setup: Vision feeds descriptions to Planner and Agent, and an extra set of calls is performed to obtain Experts. | 2403.05468v1 | 1 |
|
Figure 2:Sample screenshot fed into Vision, taken directly from room C in E1M1.
The output from the model is in Figure3. | 2403.05468v1 | 2 |
|
Figure 4:Map evaluated in this paper, corresponding to Episode 1, Map 1 (E1M1, “Hangar”) of Doom[44]. The trajectory from our walkthrough follows the sequence A, B, C, D; starting at A (leftmost checkered flag) and ending in D (rightmost checkered flag).
Depicted is a typical human run with no exploration.
The player must flip a switch in D to end the game.
Doors are denoted by an orange line.
The zig-zag room (C) contains pools of acid (not pictured) that damage the player if walked on. | 2403.05468v1 | 4 |
|
Figure 5:Trajectories by Agent. The starting point is at the leftmost checkered flag; the ending (a switch that must be activated) is the rightmost checkered flag.
Other areas (e.g., G, J) are secrets.
Time-outs are denoted by ✚ and deaths by ✖.Clockwise from the top left: naïve, with walkthrough, with k-levels, and with planner. The naïve prompt could not leave the first room and got frequently stuck.
It died twice at the hand of the zombies in room B; once due to shooting an explosive barrel.
The walkthrough prompt made it to the second room, but got stuck often in the corners and was killed by the zombies.
The planning prompt was able to kill the zombies in room B, and almost finished the map.
K-levels was consistently able to reach room C, although died by stepping on the acid.
It once managed to open a secret room, but could not backtrack and timed out. | 2403.05468v1 | 5 |
|
Figure 1:We represent robots and the task space with point clouds, and solve a trajectory optimization problem for joint motion and grasp planning. | 2403.05466v2 | 1 |
|
Figure 3:(a) A tabletop scene for grasping with a Fetch robot. (b) A 3D point cloud of the scene computed using a depth image from the camera on the robot. (c) Reaching a grasping goal can be formulated as matching 3D points on the robot gripper. (d) Visualization of the signed distance field of the task space. Cyan points are with negative distances, and yellow points are with positive distances. | 2403.05466v2 | 3 |
|
Figure 1:Illustration of the R2D2 algorithm. Both image iterates𝒙(i−1)superscript𝒙𝑖1\bm{x}^{(i-1)}bold_italic_x start_POSTSUPERSCRIPT ( italic_i - 1 ) end_POSTSUPERSCRIPTand associated residual dirty images𝒓(i−1)superscript𝒓𝑖1\bm{r}^{(i-1)}bold_italic_r start_POSTSUPERSCRIPT ( italic_i - 1 ) end_POSTSUPERSCRIPTare fed to R2D2 DNNs as input. R2D2 DNNs’ output are then used to update the next image iterates. The sequence of the image iterates and corresponding residual dirty images are indicated with dashed red and green arrows, respectively. The sequence of the learned residual images are indicated with blue arrows. | 2403.05452v3 | 1 |
|
Figure 2:Illustration of the U-Net core architecture underpinning the different incarnations of the R2D2 algorithm. The convolutional layers of the network, represented as boxes, apply multi-channel convolutions, followed by down-sampling (contracting path) or up-sampling (expanding path). The outcome of these layers are 3D feature maps, whose dimensions are specified around the corresponding boxes: the 2D spatial dimension at the bottom center, and the number of feature channels at the outer edge. | 2403.05452v3 | 2 |
|
Figure 3:Selection of raw low-dynamic range images (linear scale; top row), and corresponding ground-truth images after pre-processing (logarithmic scale; bottom row). The training dataset includes medical images (first column), and optical astronomy images (second column). The validation dataset comprises images of giant radio galaxies (e.g., Messier 87; third column), and radio galaxy clusters (e.g., Abell 746; fourth column). | 2403.05452v3 | 3 |
|
Figure 4:A selection of simulated 2D Fourier sampling patterns with the antenna configuration of the VLA (top row) and resulting dirty images (linear scale; bottom row). The depicted dirty images correspond to the ground-truth images shown in Figure3. | 2403.05452v3 | 4 |
|
Figure 5:Results of the dataset pruning procedure. Evolution of the size of the training dataset shown as a fraction of the size of the initial training dataset, throughout the iterations of R2D2 and both R3D3 realizations (R3D33LsuperscriptR3D33L\textrm{R3D3}^{\textrm{3L}}R3D3 start_POSTSUPERSCRIPT 3L end_POSTSUPERSCRIPT,R3D36LsuperscriptR3D36L\textrm{R3D3}^{\textrm{6L}}R3D3 start_POSTSUPERSCRIPT 6L end_POSTSUPERSCRIPT). | 2403.05452v3 | 5 |
|
Figure 6:Progress of the reconstruction quality (left column: SNR, middle column: logSNR) and data fidelity (right column:σ¯res.subscript¯𝜎res.{\overline{\sigma}}_{\textrm{res.}}over¯ start_ARG italic_σ end_ARG start_POSTSUBSCRIPT res. end_POSTSUBSCRIPT) across the iterations of R2D2 and both realizations of R3D3. Values of the different metrics, achieved at convergence by the benchmark algorithms uSARA, AIRI and CLEAN, are reported via horizontal lines. From top to bottom: results of Experiments I-IV. In each scenario, the reported values are averages over a test dataset composed of 100 inverse problems. | 2403.05452v3 | 6 |
|
Figure 7:Experiment I: reconstructions results of the galaxy cluster Abell 2034. The first and third rows show the ground truth image and the estimated images obtained by the imaging algorithms (logarithmic scale). The second and fourth row show the dirty image and the corresponding residual dirty images (linear scale). For R3D3, we showcase the results of its realization withJ=6𝐽6J=6italic_J = 6layers in its R2D2-Net components. Values of (SNR, logSNR) are reported below estimated images.σ¯res.subscript¯𝜎res.{\overline{\sigma}}_{\textrm{res.}}over¯ start_ARG italic_σ end_ARG start_POSTSUBSCRIPT res. end_POSTSUBSCRIPTvalues are indicated below the residual dirty images. | 2403.05452v3 | 7 |
|
Figure 8:Experiment II: reconstructions results of the radio galaxy 3C353. The first and third rows show the ground truth image and the estimated images obtained by the imaging algorithms (logarithmic scale). The second and fourth row shows the dirty image and the corresponding residual dirty images (linear scale). For R3D3, we showcase the results of its realization withJ=6𝐽6J=6italic_J = 6layers in its R2D2-Net components. Values of (SNR, logSNR) are reported below estimated images. Theσ¯res.subscript¯𝜎res.{\overline{\sigma}}_{\textrm{res.}}over¯ start_ARG italic_σ end_ARG start_POSTSUBSCRIPT res. end_POSTSUBSCRIPTvalues are indicated below the residual dirty images. | 2403.05452v3 | 8 |
|
Figure 9:Experiment III: reconstructions results of the galaxy cluster PSZ2 G165.68+44.01. The first and third rows show the ground truth image and the estimated images obtained by the imaging algorithms (logarithmic scale). The second and fourth row shows the dirty image and the corresponding residual dirty images (linear scale). For R3D3, we showcase the results of its realization withJ=6𝐽6J=6italic_J = 6layers in its R2D2-Net components. Values of (SNR, logSNR) are reported below estimated images. Theσ¯res.subscript¯𝜎res.{\overline{\sigma}}_{\textrm{res.}}over¯ start_ARG italic_σ end_ARG start_POSTSUBSCRIPT res. end_POSTSUBSCRIPTvalues are indicated below the residual dirty images. | 2403.05452v3 | 9 |
|
Figure 10:Experiment IV: reconstructions results of the radio galaxy Messier 106. The first and third rows show the ground truth image and the estimated images obtained by the imaging algorithms (logarithmic scale). The second and fourth row shows the dirty image and the corresponding residual dirty images (linear scale). For R3D3, we showcase the results of its realization withJ=6𝐽6J=6italic_J = 6layers in its R2D2-Net components. Values of (SNR, logSNR) are reported below estimated images. Theσ¯res.subscript¯𝜎res.{\overline{\sigma}}_{\textrm{res.}}over¯ start_ARG italic_σ end_ARG start_POSTSUBSCRIPT res. end_POSTSUBSCRIPTvalues are indicated below the residual dirty images. | 2403.05452v3 | 10 |
|
Figure 1:Visualization of images (a), raw feature maps (b), and refined feature maps (c). Channel and spatial attention is applied to raw features, emphasizing on the important regions and making them valuable distillation source. | 2403.05451v2 | 1 |
|
Figure 7:Visual representation of the performance of proposed method in terms of per-class mIoU using ResNet18 network on PascalVoc validation set. | 2403.05451v2 | 7 |
|
Figure 8:Comparison of mIoU per class among LAD, DistKD, and AttnFD on Cityscapes validation set, employing a ResNet18 backbone for the student network. | 2403.05451v2 | 8 |
|
Figure 1:The unprecedented increase of volatility in the (German) intraday electricity market is illustrated by the volume-weighted average price (IDFull) for all hourly products for 2021 and 2022 shown as a thin line. The thick line is a weekly rolling average. The data was provided byEPEX Spot SE. | 2403.05441v3 | 1 |
|
Figure 2:Illustration of the transformation proposed inKulakov and Ziel (2019)to obtain the slope of the merit-order curve as a measure for elasticityη𝜂\etaitalic_ηin the DA market. The left side shows aggregated supply (SEP) and demand (DEM) curves as obtained fromEPEX Spot SEfor the delivery hour 7:00 - 8:00 am on 4 November 2022, with the DA pricePdasubscript𝑃daP_{\mathrm{da}}italic_P start_POSTSUBSCRIPT roman_da end_POSTSUBSCRIPTpublished in the European trading zone as a horizontal line. The right side depicts these curves after the transformation, in which the elasticity from the demand side is transferred to the supply side. Exemplary slopes are indicated by straight lines for two volumes marked by closed circles. A finite difference of 500 MWh has been used to determine the slopes. | 2403.05441v3 | 2 |
|
Figure 3:Live CID price and volume indices and statistics calculated from transaction data provided byEPEX Spot SEfor the exemplary delivery hour 7:00 - 8:00 am CET on 4 November 2022 as a function of forecast creation timeτ𝜏\tauitalic_τ. The horizontal dashed lines indicate published values by EPEX. The dashed vertical line marks the beginning of delivery dayd𝑑ditalic_d, i.e. hours left from this line are on dayd−1𝑑1d-1italic_d - 1and on the right on dayd𝑑ditalic_d. Solid vertical lines indicate 3 h before delivery start and 1 h before delivery start where applicable, and delivery start itself. | 2403.05441v3 | 3 |
|
Figure 4:Measures of elasticity of the DA market and the CID market using spot price and live IDFull for the exemplary delivery hour 7:00 - 8:00 am CET on 4 November 2022. Here the three finite differencesΔpΔ𝑝\Delta proman_Δ italic_pare used that are also used in the data model described in Section3. The underlying live IDFull trajectory is shown in Figure3, details on the elasticity estimation is illustrated in Figure2. | 2403.05441v3 | 4 |
|
Figure 8:Mean absolute errors (MAEs) for scenarios (a) - (d), where forecast creation time is fixed, and all following delivery hours are forecasted. The point estimatey^^𝑦\hat{y}over^ start_ARG italic_y end_ARGfor the end-of-day IDFull is given by (13), the live IDFullPidfullsubscript𝑃idfullP_{\mathrm{idfull}}italic_P start_POSTSUBSCRIPT roman_idfull end_POSTSUBSCRIPTis used as reference. | 2403.05441v3 | 8 |
|
Figure 9:The top left chart depicts the difference betweeny^^𝑦\hat{y}over^ start_ARG italic_y end_ARGand the live IDFull in terms of their MAE for scenarios (a) - (d), summarizing Figure8. The bottom shows the same for scenarios (e) using OMP and (f) using LASSO for feature selection. Results below the solid red line possess smaller MAEs than their live IDFull references. In the top right chart, the MAE values are averaged across hours, with OMP showing an average decrease of5.9%percent5.95.9\,\%5.9 %in absolute errors compared to the live benchmark and a22.7%percent22.722.7\,\%22.7 %reduction compared to LASSO, while LASSO leads to an28.8%percent28.828.8\,\%28.8 %increase of absolute errors compared to the live benchmark. | 2403.05441v3 | 9 |
|
Figure 11:The accuracy of the estimators (16) and (17) for the sign of thespreadand therestare shown, together with the IDFull reference for the spread sign, cf. Figure10. Here, the scenarios (e) and (f) investigate the difference in performance using OMP and LASSO for feature selection. OMP demonstrates an average increase of1.7%percent1.71.7\,\%1.7 %in spread sign forecast accuracy compared to the live benchmark, and increases of12.1%percent12.112.1\,\%12.1 %and10.0%percent10.010.0\,\%10.0 %for spread and rest sign forecast accuracy compared to LASSO. LASSO leads to a decrease in accuracy by−9.0%percent9.0-9.0\,\%- 9.0 %compared to the live benchmark. The accuracy is determined as number of correctly forecasted days across all hours over the total number of forecasts. | 2403.05441v3 | 11 |
|
Figure 12:The Empirical Coverage (left) and the resulting Average Coverage Error (ACE) (right) are shown for the fixed creation timesτ𝜏\tauitalic_τof scenarios (a) - (d). The solid green lines represent the theoretical expectation in case of perfect coverage. | 2403.05441v3 | 12 |
|
Figure 13:The Empirical Coverage (top panel) and the resulting Average Coverage Error (ACE) (bottom panel) are shown for scenarios (e) and (f) with fixed lag between forecast creation and delivery begin. The left panel uses OMP feature selection, the right panel LASSO feature selection. The solid green lines represent the theoretical expectation in case of perfect coverage. | 2403.05441v3 | 13 |
|
Figure 14:Comparison between the use of OMP and LASSO for feature selection for scenarios (e) and (f) averaged across all days and hours of the test period. On the left, we show the Average Coverage Error (ACE), with OMP showing an average34.59%percent34.5934.59\,\%34.59 %decrease compared to LASSO. On the right, we show the Continuous Ranked Probability Score (CRPS) defined in (18), with OMP showing an average20.21%percent20.2120.21\,\%20.21 %decrease compared to LASSO. | 2403.05441v3 | 14 |
|
Figure 1:Illustration of the large variability of item-item cosine similaritiescosSim(B,B)𝑐𝑜𝑠𝑆𝑖𝑚𝐵𝐵cosSim(B,B)italic_c italic_o italic_s italic_S italic_i italic_m ( italic_B , italic_B )on the same data due to different modeling choices. Left: ground-truth clusters (items are sorted by cluster assignment, and within each cluster by descending baseline popularity). After training w.r.t. Eq.1, which allows for arbitrary re-scaling of the singular vectors inVksubscript𝑉𝑘V_{k}italic_V start_POSTSUBSCRIPT italic_k end_POSTSUBSCRIPT, the center three plots show three particular choices of re-scaling, as indicated above each plot. Right: based on (unique)B𝐵Bitalic_Bobtained when training w.r.t. Eq.2. | 2403.05440v1 | 1 |
|
Figure 2:VideoElevator for improved text-to-video generation.Top:Taking textτ𝜏\tauitalic_τas input, conventional T2V performs both temporal and spatial modeling and accumulates low-quality contents throughout sampling chain.Bottom:VideoElevator explicitly decompose each step intotemporal motion refiningandspatial quality elevating, where the former encapsulates T2V to enhance temporal consistency and the latter harnesses T2I to provide more faithful details,e.g.,dressed in suit.
Empirically, applying T2V in several timesteps is sufficient to ensure temporal consistency. | 2403.05438v1 | 2 |
|
Figure 4:Qualitative results enhanced with foundational T2I.As one can see, VideoElevator manages to enhance the performance of T2V baselines with StableDiffusion V1.5 or V2.1-base, in terms of frame quality and text alignment.
For frame quality, the videos enhanced by VideoElevator contain more details than original videos.
For text alignment, VideoElevator also produces videos that adhere better to prompts, where inconsistent parts of baselines are colored inorange.Please watch videos in website for better view. | 2403.05438v1 | 4 |
|
Figure 5:Qualitative results enhanced with personalized T2I.With the power of personalized T2I, VideoElevator enables ZeroScope and LaVie to produce various styles of high-quality videos.
Compared to personalized AnimateDiff, VideoElevator captures more faithful styles and photo-realistic details from personalized T2I,e.g.,sunset time lapse at the beach.Please watch videos in website for better view. | 2403.05438v1 | 5 |
|
Figure 6:Qualitative results using personalized SDXL.We select AnimateDiff XL as T2V baseline and leverage two types of personalized SDXL to enhance it, including(a)DynaVision finetunes SDXL based on LoRA and slightly updates its parameters.(b)Diffusion-DPO largely updates the parameters of SDXL. | 2403.05438v1 | 6 |
|
Figure 7:Qualitative ablation on low-pass frequency filter (LPFF).(a)Not using LPFF leads to incoherence in appearance (highlighted inred circle).(b)Using temporal LPFF visibly improves temporal consistency without hurting frame quality.(c)Adding both spatial and temporal LPFF makes video frames blurry.Please watch videos in website for better view. | 2403.05438v1 | 7 |
|
Figure 1:Illustration of SAM’s ambiguity problem.
The ground truth is circled by a red dashed circle, while the predicted mask is depicted by a yellow solid line. | 2403.05433v1 | 1 |
|
Figure 4:Illustration ofP2SAMsuperscriptP2SAM\text{P}^{2}\text{SAM}P start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT SAM’s improvement. A blue star represents the selected positive-point prompt. | 2403.05433v1 | 3 |
|
Figure 1:Example vehicle detection before and after the adversarial attack in the altitudes of 50m, 90m in the proposed altitude-sensitive EVD4UAV dataset. Due to the space limit, the altitude of 70m is skipped. | 2403.05422v2 | 1 |
|
Figure 2:Total number of (a) images and (b) fine-grained annotated vehicles of our EVD4UAV in different UAV altitudes (50m, 70m, 90m). | 2403.05422v2 | 2 |
|
Figure 3:Example vehicle images in EVD4UAV with labeled horizontal bounding boxes, rotated bounding boxes, and instance-level segmentation masks from top to bottom. | 2403.05422v2 | 3 |
|
Figure 4:Proportions of (a) images in four scenes, (b) annotated vehicle color attribute, (c) annotated vehicle type attribute in EVD4UAV. | 2403.05422v2 | 4 |
|
Figure 5:Schematic representation of adversarial patch generation and application processes for white-box and black-box attacks on UAV vehicle detectors, highlighting the path from patch creation to detector evasion and classification discrepancy. | 2403.05422v2 | 5 |
|
Figure 6:The detect bounding boxes and confidence results predicted by different models. The\textBBP\textCLIP\text𝐵𝐵subscript𝑃\text𝐶𝐿𝐼𝑃\text{BBP}_{\text}{CLIP}italic_B italic_B italic_P start_POSTSUBSCRIPT end_POSTSUBSCRIPT italic_C italic_L italic_I italic_Pand\textBBP\textD\text𝐵𝐵subscript𝑃\text𝐷\text{BBP}_{\text}{D}italic_B italic_B italic_P start_POSTSUBSCRIPT end_POSTSUBSCRIPT italic_Dare detected by YOLOv8 detector. Other attack methods are targeting their own detectors. | 2403.05422v2 | 6 |
|
Figure 7:Example images of attack patches and confidence score on different altitudes. The detection results of\textBBP\textCLIP\text𝐵𝐵subscript𝑃\text𝐶𝐿𝐼𝑃\text{BBP}_{\text}{CLIP}italic_B italic_B italic_P start_POSTSUBSCRIPT end_POSTSUBSCRIPT italic_C italic_L italic_I italic_P,\textBBP\textD\text𝐵𝐵subscript𝑃\text𝐷\text{BBP}_{\text}{D}italic_B italic_B italic_P start_POSTSUBSCRIPT end_POSTSUBSCRIPT italic_D, Naiver Cover, and Random Noise are detected by YOLOv8 detector. Other attack methods are targeting their own detectors. | 2403.05422v2 | 7 |
|
Figure 2:The illustration of upsample block used in SatMAE++ framework. Input featuresX𝑋Xitalic_Xare upsampled by utilizing the transpose convolution operation. Afterwards, a residual block which is composed of two convolution layers is employed to enhance the upsampled features given asX~~𝑋\tilde{X}over~ start_ARG italic_X end_ARG. | 2403.05419v1 | 2 |
|
Figure 3:SatMAE++ reconstruction results at multi-scale level. Examples from fMoW-Sentinel dataset are shown here. For illustration, we show the RGB channels of the multi-spectral data here. The images are reconstructed at resolutions of(H,W)𝐻𝑊(H,W)( italic_H , italic_W ),(2H,2W)2𝐻2𝑊(2H,2W)( 2 italic_H , 2 italic_W ), and(4H,4W)4𝐻4𝑊(4H,4W)( 4 italic_H , 4 italic_W ), respectively. We observe that the proposed model provide better reconstruction results compared to SatMAE at resolution of(H,W)𝐻𝑊(H,W)( italic_H , italic_W ). | 2403.05419v1 | 3 |
|
Figure 5:Illustration of finetuning convergence on validation set of fMoW-Sentinel dataset. We observe that the model pre-trained with multi-scales achieves faster convergence as compared the model pre-trained with single or less scales. The model trained with single scale achieves highest score of 61.61 at 20th epoch whereas the model that utilised three scales in pre-training converges earlier and achieves highest score at 12th epoch. | 2403.05419v1 | 5 |
|
Figure 1:Self-supervised strategy based on negative augmentation. To explore target correspondence in SIRST, we propose a negative augmentation approach to generate massive negatives for self-supervised representation learning. | 2403.05416v1 | 1 |
|
Figure 2:The quality, quantity, and diversity of the infrared data have a significant impact on the detection of small targets. By applying a self-supervised learning paradigm with massive pseudo-data, our negative generation strategy has achievedfaster convergence rate, less training loss and better meanIoU𝐼𝑜𝑈IoUitalic_I italic_o italic_U. | 2403.05416v1 | 2 |
|
Figure 3:The qualitative results of different SIRST detection methods. The correctly detected targets, false alarms, and weak detection regions are highlighted usingred,yellow, andgreendashed circles respectively. Our method performs accurate target localization with a lower false alarm rate. | 2403.05416v1 | 3 |
|
Figure 4:Illustration of the proposed massive negatives synthesis framework. (a) Noise Sampling and Modeling module. The input image is first divided into equally sized local regions. Qualified noise sampling regions are selected and resized to fetch real-world noise. Then, the training samples are mixed with real-world noise; (b) Negative Augmentation module. Among the generated samples, the challenging small targets are further processed with negative augmentation to produce massive negatives; (c) Self-supervised Learning. By utilizing these negative samples and their corresponding labels, we can implement self-supervised representation learning to learn richer feature representations. | 2403.05416v1 | 4 |
|
Figure 5:A demonstration of noise sampling in infrared small target detection dataset. High-variance noise can affect the model’s ability to recognize small targets. The yellow dashed circle highlights the introduced texture. By selecting Noise-prone Region (Anoisesubscript𝐴𝑛𝑜𝑖𝑠𝑒A_{noise}italic_A start_POSTSUBSCRIPT italic_n italic_o italic_i italic_s italic_e end_POSTSUBSCRIPT), our framework can fetch diverse noise in infrared sensors to facilitate realistic sample augmentation in SIRST-5K. | 2403.05416v1 | 5 |
|
Figure 6:With the negative augmentation, limited infrared images are renewed with richer diversity. To make the discriminative model learn effective features on small targets, a negative augmentation strategy is developed for the special yet challenging target augmentation. | 2403.05416v1 | 6 |
|
Figure 7:Visualization of the samples in SIRST-5K. Among our framework, Real-world Noise2Noise Displacement and Negative Augmentation methods adjust the noise weight and rotation (angle) of the generated image for self-supervised learning. | 2403.05416v1 | 7 |
|
Figure 8:Attention Feature Fusion module is used to reduce the semantic gap during the multi-layer feature fusion stage. This diagram representsChannel Attention Module (left)andSpatial Attention Module (right), respectively | 2403.05416v1 | 8 |
|
Figure 9:The qualitative results obtained from different SIRST detection methods are shown. For better visualization, the small target area is enlarged. Our method performs accurate target localization with a lower false alarm rate. | 2403.05416v1 | 9 |
|
Figure 10:Visualization curves of ablation studies for the real-world Noise2Noise Displacement module and Negative Augmentation modules. | 2403.05416v1 | 10 |
|
Figure 1:The illustration of FedSAM and FedMSA. FedSAM is a federated learning framework with (a) SAM. FedMSA is a federated learning framework with (b) MSA. | 2403.05408v2 | 1 |
|
Figure 2:The CCI values for identified regions based on NF-iVAE and iVAE. a. results for the case that attention is confounder b. results for the case that visual is confounder c. the relation between dorsal, ventral and visual regions | 2403.05407v2 | 2 |
|
Figure 3:Probability of Visual Cortex Nodes Ranked Among Top 5 by CCI Across 30 Runs. The y-axis denotes the frequency of inclusion in the top 5 rankings, and the x-axis lists the nodes. | 2403.05407v2 | 3 |
|
Figure 2:Graphical illustration of different operations of the HTV-Trans: (a) generative process of HTPGM, (b) the inference scheme of HTPGM. (c) the fusion of different scale information and stationarization input series for forecasting. | 2403.05406v1 | 2 |
|
Figure 5:The effectiveness evaluation of hierarchical architecture on ETTh1(left), ETTh2(middle) and ETTm1(right) dataset. | 2403.05406v1 | 5 |
|
Figure 6:The effectiveness of parameterα𝛼\alphaitalic_αon dataset ETTm1 with different predict length. | 2403.05406v1 | 6 |
|
Figure 3:Overview of DualBEV:Initially, we employ SceneNet to predict the depthD𝐷Ditalic_D(Projection probability) and instance maskM𝑀Mitalic_M(Image probability) of input images. Subsequently, the Prob-LSS stream follows the BEVPoolv2[5]to generate LSS feature. Concurrently, the HeightTrans stream utilizes the Prob-Sampling to project pre-defined 3D points onto the 2D space, retrieving corresponding image features. Throughout this process, all features are accompanied by probabilities derived from the depth map and instance mask. Finally, we fuse two streams and predict the BEV probabilityP𝑃Pitalic_Pby leveraging the DFF module, resulting in the final BEV featureF𝐹Fitalic_F. | 2403.05402v2 | 6 |
|
Figure 5:Dual Feature Fusion Module: Dual features are first concatenated and then passed into the CAF module for fusion. Subsequently, the SAE-ProbNet is utilized to obtain the BEV probability for the final BEV feature. | 2403.05402v2 | 11 |
|
Figure 2:Qualitative analysis of the proposed HistGen model. Words highlighted in bold green indicate alignment between our model’s generated results and the ground truth. Conversely, words underlined in orange represent diagnostic details that our model fails to capture. The first two examples highlight the superior captioning capability of our model, with it accurately diagnosing provided WSIs. The diagnoses closely align with the ground truths, differing only in minor, non-critical aspects. In the third example, our model successfully makes the correct prediction, despite the absence of the detailed context present in the ground truth. | 2403.05396v2 | 2 |
|
Figure 3:WSI distribution for DINOv2 ViT-L feature extractor pre-training. We have collected over 30 different pathology datasets containing over 60 primary sites. Patches are extracted from whole slide images at level00, with dimensions of512×512512512512\times 512512 × 512. These patches are subsequently resized to224×224224224224\times 224224 × 224for pre-training the feature extractor. This figure shows the details of our collected WSIs. | 2403.05396v2 | 3 |
|
((a))Empirical probability over 50 runs that (4) is verified as a function ofm𝑚mitalic_mandk𝑘kitalic_k. | 2403.05395v1 | 1 |
|
((a))Empirical probability over 50 runs that (4) is verified as a function ofm𝑚mitalic_mandk𝑘kitalic_k. | 2403.05395v1 | 2 |
Subsets and Splits