image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 9:Trajectories (top row) and distance to equilibrium over synchronization rounds (bottom row) of GDA (K=1𝐾1K=1italic_K = 1) and Decoupled SGDA withK={2,5}𝐾25K=\{2,5\}italic_K = { 2 , 5 }on the (toyGAN) problem (d=2𝑑2d=2italic_d = 2).𝐂𝐂\mathbf{C}bold_Cin (QG) is a constant here—the larger, the stronger the interactive term.Left-to-right:decreasing the constantc∈{10,3.5,2,7,0}𝑐103.5270c\in\{10,3.5,2,7,0\}italic_c ∈ { 10 , 3.5 , 2 , 7 , 0 }. | 2501.14652v1 | 9 |
|
Figure 10:Comparison of Decoupled SGDA and Local SGDA under different noise settings.Each plot shows the smallest gradient norm achieved by both algorithms over 100 communication rounds, with varying interaction levels and noise variances.Top Row:Different settings of noise variances in off-diagonal entries (interaction noise).Left to Right:Increasing values of the constantc𝑐citalic_ccontrolling the interactive term’s strength in the game. Decoupled SGDA consistently outperforms Local SGDA in scenarios where off-diagonal noise is significant, achieving lower gradient norms with fewer communication rounds. | 2501.14652v1 | 10 |
|
Figure 6:An illustration of a dense and sparsepDsubscript𝑝𝐷p_{D}italic_p start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT. In the case of the sparse distribution, annealing would have a lower impact as the newly initialised policy would start out receiving much higher rewards. | 2501.14856v1 | 8 |
|
Figure 7:Discriminator variance experiments. The policy was first trained for2e62𝑒62e62 italic_e 6,5e65𝑒65e65 italic_e 6, and10e610𝑒610e610 italic_e 6data samples. Then, with the policy updates paused, discriminator training was continued. We plot the rewards received across several independent policy rollouts and observe a high variance. Given that the policy is unchanging, a high variance in the reward indicates poor reinforcement learning. | 2501.14856v1 | 9 |
|
Figure 8:Discriminator non-smoothness experiments. Plots show the mean and std. discriminator prediction over a large batch of perturbed expert data samples at varying levels of perturbation (where a distance of 0 on the x-axis corresponds to unperturbed expert data inpDsubscript𝑝𝐷p_{D}italic_p start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT). Notice the high value of the std. compared to the mean at any given distance frompDsubscript𝑝𝐷p_{D}italic_p start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT. | 2501.14856v1 | 10 |
|
Figure 9:Perfect discriminator experiments on the walking task (multi-clip dataset with 74 motions clips). The policy was first trained for0.5e60.5𝑒60.5e60.5 italic_e 6,2e62𝑒62e62 italic_e 6, and5e65𝑒65e65 italic_e 6data samples. Then, with the policy updates paused, the discriminator was retrained. We find that the discriminator very quickly learns to perfectly distinguish betweenpDsubscript𝑝𝐷p_{D}italic_p start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPTandpGsubscript𝑝𝐺p_{G}italic_p start_POSTSUBSCRIPT italic_G end_POSTSUBSCRIPT(notice the logarithmic scale). | 2501.14856v1 | 11 |
|
Figure 1:An illustration of the sample and task construction of the DEDC framework. An example of the sample construction is shown on the left. On the right is an illustration of the task construction when the sample on the left is used as the test sample. | 2501.14649v1 | 1 |
|
Figure 3:An illustration of the compositional gap between the test sample and its demonstration samples. A and B indicate the different output types on the edges. | 2501.14649v1 | 3 |
|
Figure 4:An illustration of the two types of settings for counter-intuitive symbolic names. The arrow indicates that a primitive with the meaning on the left end uses the name on the right end. | 2501.14649v1 | 4 |
|
Figure 2:Illustration of SyncAnimation Framework: Given a single image and audio input, the preprocessing stage extracts 3DMM parameters in NeRF space as references for Audio2Pose and Audio2Emotion (or alternatively, noise). The framework then progressively generates the upper body, head, and final lip refinement. Audio2Pose ensures poses consistency with the audio for upper-body generation, while Audio2Emotion aligns facial expression with the audio for head rendering, in conjunction with the generated poses. | 2501.14646v1 | 2 |
|
Figure 3:Audio2Pose is designed to reconstruct stable head pose offsets (𝐎𝐟𝐟(𝐞\mathbf{Off}\text{ }(\mathbf{e}bold_Off ( bold_e) and𝐎𝐟𝐟(𝐭)𝐎𝐟𝐟𝐭\mathbf{Off}\text{ }(\mathbf{t})bold_Off ( bold_t )) using audio and monocular input. Where, Wav2attr, a pre-trained audio encoder, is employed to encode audio vectors containing character-specific information. Additionally, a Gaussian-based VAE is integrated to introduce a diversity template𝐒posesubscript𝐒pose\mathbf{S}_{\text{pose}}bold_S start_POSTSUBSCRIPT pose end_POSTSUBSCRIPT, while a stability model𝐃posesubscript𝐃pose\mathbf{D}_{\text{pose}}bold_D start_POSTSUBSCRIPT pose end_POSTSUBSCRIPTis implemented based on the poses labels with high dropout rate (DP=0.6DP0.6\text{DP}=0.6DP = 0.6), improving the effectiveness of pose reconstruction. | 2501.14646v1 | 3 |
|
Figure 4:Audio2Emotion is designed to learn and reconstruct 3DMM expression offsets (𝐎𝐟𝐟(𝐛)𝐎𝐟𝐟𝐛\mathbf{Off}\text{ }(\mathbf{b})bold_Off ( bold_b )) using audio and monocular input. The structure is similar to that of Audio2Pose, but due to the weak correlation between facial expressions and audio, and the periodic nature of blinking, we modify the diversity template𝐒expsubscript𝐒exp\mathbf{S}_{\text{exp}}bold_S start_POSTSUBSCRIPT exp end_POSTSUBSCRIPT. It is replaced by a conditional VAE guided by periodic time features𝐓τsubscript𝐓𝜏\mathbf{T}_{\tau}bold_T start_POSTSUBSCRIPT italic_τ end_POSTSUBSCRIPTand context-dependent audio features𝐀seqsubscript𝐀seq\mathbf{A}_{\text{seq}}bold_A start_POSTSUBSCRIPT seq end_POSTSUBSCRIPT. | 2501.14646v1 | 4 |
|
Figure 5:Visual comparison with outputs of baselines. GAN-based and NeRF-based methods generate avatar with fixed poses and expressions, while SD-based methods introduce expression changes but lack face detail and pose movement. SyncAnimation uniquely achieves jointly generative, audio-driven realistic expressions and movable poses. | 2501.14646v1 | 5 |
|
Figure 7:Ablation Study on𝐃expsubscript𝐃exp\mathbf{D_{\text{exp}}}bold_D start_POSTSUBSCRIPT exp end_POSTSUBSCRIPTand𝐒expsubscript𝐒exp\mathbf{S_{\text{exp}}}bold_S start_POSTSUBSCRIPT exp end_POSTSUBSCRIPTin OOD audio inference. Removing them will result in (a) and (b). | 2501.14646v1 | 7 |
|
Figure 2:Visual example of PPM-Reg in a shape matching experiment using Cramer or MMD as the main loss function. 1st row: Plots of a reference point cloud (in blue) and the initial condition of a random point cloud (in orange). 2nd Row: Plots of 2-Wasserstein distance between 1-dimensional persistent homology between the reference shape and training shape over optimization steps. | 2501.14641v1 | 2 |
|
Figure 3:CMMD (a,d), FDDinov2(b,e) and WDlatent(c,f) versus training epochs for the AnimeFace (a-c) and CelebA (d-f) dataset. 10K samples are randomly generated to compute distances; moving averages with a window of 5 are used to smooth the values. Distances recorded every 160 epochs. | 2501.14641v1 | 3 |
|
Figure 4:Plots of the shape matching experiment at convergence after 16000 steps. (a) and (e) use only the Cramer loss. (b) and (f) use Cramer + PPM-Reg. (c) and (g) use only the MMD loss. (d) and (h) use MMD + PPM-Reg. | 2501.14641v1 | 4 |
|
Figure 6:CMMD (a,e), FDDinov2(b,f) and WDlatent(c,g) versus training epochs for the CelebA (a-c) and LSUN Kitchen (d-f) dataset for the64×64646464\times 6464 × 64image generation. 10K samples are randomly generated to compute the distances, and moving averages with a window of 5 are used to smooth the values. For CelebA, distances are recorded every 160 epochs. For LSUN Kitchen, distances are recorded every 20 epochs. | 2501.14641v1 | 8 |
|
Figure 7:Generated images from SSL experiment from trainedg𝝎subscript𝑔𝝎g_{\boldsymbol{\omega}}italic_g start_POSTSUBSCRIPT bold_italic_ω end_POSTSUBSCRIPTfor MNIST using (a) Cramer Distance and (b) PPM-Reg. | 2501.14641v1 | 9 |
|
Figure 8:Network architecture of generator (a), discriminator (b) for32×32323232\times 3232 × 32unconditional image generation. | 2501.14641v1 | 10 |
|
Figure 9:Network architecture of generator (a), discriminator (b) for64×64646464\times 6464 × 64unconditional image generation. | 2501.14641v1 | 11 |
|
Figure 1:Effect of scaling the LLM judge and increasing the number of instructions on Spearman correlation. In contrast to human agreement, neither Alpaca-Eval, Arena-Hard, nor their union distinguishes the quality difference between 32B and 72B models. | 2501.17178v2 | 1 |
|
Figure 4:Illustration of the selection process.
All 4480 configurations are first evaluated on 400 instructions (left), the top 1200 configurations are then evaluated on 1200 instructions (center) and finally the top 400 configurations are evaluated on 3548 instructions (right). The color denotes the ranking assigned by the non-dominated sort procedure. | 2501.17178v2 | 4 |
|
Figure 5:We plot the cost per annotation and human agreement of all 4480 judges when using 400 instructions. The model family and the number of parameters are represented with color and size respectively. | 2501.17178v2 | 5 |
|
Figure 6:Fraction of time each hyperparameter appears in the top 100 configurations for small (<10B) and large models (>10B). | 2501.17178v2 | 6 |
|
Figure 7:Prompt performance stability across different models. We show the correlation matrix between models when looking at their performance on all the 80 different prompts. | 2501.17178v2 | 7 |
|
Figure 11:Illustration of the non-dominated sorting approach. The process first computes the Pareto front, assigning top ranks to the points in this layer. Next, the Pareto front is determined for the remaining points, which are then assigned the next set of rankings. This process continues iteratively until all points are ranked. To resolve ties within each layer, a heuristic is applied to balance both sparsity and coverage. | 2501.17178v2 | 11 |
|
Figure 12:Scatter plot of cost and human agreement on the 400 validation instructions for all judges. We color-code each hyperparameter differently to illustrate the performance of all judges. Even using the same LLM model, there is a large spread of performance when varying other hyperparameters without an obvious pattern to distinguish the best configuration which motivates the need to search for optimal hyperparameters. | 2501.17178v2 | 12 |
|
Figure 13:Correlation for different fidelity sizes.
For each fidelity, we randomly split the instructions into two buckets and plot the human-agreement on the first bucket versus the same metric computed on the second bucket of instructions for all the available judges, we also report the Spearman correlationρ𝜌\rhoitalic_ρbetween the two groups. This allows to see the correlation one can obtain between the two sets. For the final fidelity, we measure the validation performance on 3548 instructions and use 3000 test instructions so we expect a higher correlation between validation and test scores. | 2501.17178v2 | 13 |
|
Figure 6:Average relative reconstruction error norms for the testing data for the nonlinear PAIR reconstruction and the direct encoder/decoder inversion network, for various numbers of supervised training pairs. | 2501.14636v1 | 6 |
|
Figure 8:PAIR metrics for indicating out-of-distribution. In-distribution samples from the MNIST test set are shown in dark blue, while OOD notMNIST samples are shown in light orange (5000 samples each). | 2501.14636v1 | 8 |
|
Figure 9:Scatterplot of samples from within distribution (MNIST) and outside of distribution (notMNIST) with the two most informative metrics explored earlier: the relative differences in the latent data and parameter spaces. | 2501.14636v1 | 9 |
|
Figure 1:Illustration of Wasserstein barycenters computed using different methods. The goal is to compute the barycenter of four uniform densities supported on the square, circle, heart, and cross, respectively, as displayed in the top left image. The blended shape shown in the top middle image is the barycentric density computed using our method. Barycentric densities computed using CWB and DSB with regularization strength parameterreg=0.005reg0.005\text{reg}=0.005reg = 0.005, and their thresholded versions are shown in the top right image and the bottom three images. | 2501.14635v1 | 1 |
|
Figure 2:Top row displays three exemplary digit 8 images. Bottom row displays barycenters computed by different methods using 300 iterations. | 2501.14635v1 | 2 |
|
Figure 2:Strategic Implementation Pathways forHydrogenEngines, showing primary and secondary strategies with tactical implementations. Dashed lines indicate cross-strategy synergies. | 2501.14634v1 | 2 |
|
Figure 3:Strategic Implementation Pathways for Commodore vs. Apple, showing a primary focus on core capability development and a secondary focus on resource optimization. Dashed lines indicate cross-strategy synergies. | 2501.14634v1 | 3 |
|
Figure 1:Schematic of the whole approach from the encoding scheme
to pool of refined local search functions. From this pool we
take the top-performing functions for the evaluation on test
instances. | 2501.14630v2 | 1 |
|
Figure 3:Local search function scores on the training instances
and test instances.
Each marker represents one local search function. | 2501.14630v2 | 6 |
|
Figure 2:Average efficiency of the auction run under the three scenarios fromTable1. Anchoring on our DNF learning proxy, we see that the drop-in LLM proxyωvd1subscript𝜔vd1\omega_{\textsc{vd1}}italic_ω start_POSTSUBSCRIPT vd1 end_POSTSUBSCRIPTperforms similarly, also reaching 75% efficiency in around10101010person–proxy interactions. The proxyωvd2subscript𝜔vd2\omega_{\textsc{vd2}}italic_ω start_POSTSUBSCRIPT vd2 end_POSTSUBSCRIPTperforms significantly better, reaching 75% efficiency in four interactions; proxyωnvdsubscript𝜔nvd\omega_{\textsc{nvd}}italic_ω start_POSTSUBSCRIPT nvd end_POSTSUBSCRIPT, which additionally uses a natural language question, reaches 75% efficiency in two interactions. | 2501.14625v1 | 2 |
|
Figure 3:Performance of the hybrid proxyωhsubscript𝜔h\omega_{\textsc{h}}italic_ω start_POSTSUBSCRIPT h end_POSTSUBSCRIPTand the DNF learning proxyωxorsubscript𝜔xor\omega_{\textsc{xor}}italic_ω start_POSTSUBSCRIPT xor end_POSTSUBSCRIPTwith respect to number of person–proxy interactions. Panel (a) shows thatωhsubscript𝜔h\omega_{\textsc{h}}italic_ω start_POSTSUBSCRIPT h end_POSTSUBSCRIPTachieves a good approximation of its person’s valuation significantly quicker thanωxorsubscript𝜔xor\omega_{\textsc{xor}}italic_ω start_POSTSUBSCRIPT xor end_POSTSUBSCRIPT, while still converging to the person’s exact valuation. Panel (b) shows a rapid efficiency increase initially forωhsubscript𝜔h\omega_{\textsc{h}}italic_ω start_POSTSUBSCRIPT h end_POSTSUBSCRIPT, and similar long-run efficiency. | 2501.14625v1 | 3 |
|
Figure 4:Comparison between GPT-based and other LLM models models in person robustness tests. Gpt-4o-mini and other LLM-models give similar values in response to value queries evidenced by the dots, each of which representing a single bundle, falling near the identity gpt-4o-mini value ($) = Other LLM value line ($). | 2501.14625v1 | 4 |
|
Figure 5:Preference variability and shape after iteratively adding items to an initially empty bundle, and then iteratively removing items. Each box plot displays the distribution of values obtained from with 10 iterations of the process. | 2501.14625v1 | 5 |
|
Figure 1:Densities of cosine similarity (x-axis) of context embeddingse1subscript𝑒1e_{1}italic_e start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTande2subscript𝑒2e_{2}italic_e start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTvs median similarity rating (y-axis). Note that the x-axis does not start at 0. | 2501.14617v1 | 1 |
|
Figure 1:Simulated data.A.Parametersa𝑎aitalic_a,b𝑏bitalic_b,v0subscript𝑣0v_{0}italic_v start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTandΔuΔ𝑢\Delta uroman_Δ italic_ufor simulated data with each neuron colored according to the corresponding group. Parametersa𝑎aitalic_aandb𝑏bitalic_bdo not result in separation between the groups, whereasv0subscript𝑣0v_{0}italic_v start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTandΔuΔ𝑢\Delta uroman_Δ italic_uperfectly separate the three groups.B.Four (unnormalized) randomly chosen traces (σt=1subscript𝜎𝑡1\sigma_{t}=1italic_σ start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT = 1) for each group: RS (blue), IB (orange), and CH (green). | 2501.14615v2 | 1 |
|
Figure 2:Conceptual representation of the autoregressive variational autoencoder model.A.Probabilistic model and variational posterior model in plate notation (shaded circles are observed variables).B.Architecture of the implemented autoregressive variational autoencoder (AVAE). The encoder consists of a long short-term memory (LSTM) layer followed by two fully connected linear layers in parallel, encoding the parametersμXsubscript𝜇𝑋\mu_{X}italic_μ start_POSTSUBSCRIPT italic_X end_POSTSUBSCRIPTandlogσX2superscriptsubscript𝜎𝑋2\log\sigma_{X}^{2}roman_log italic_σ start_POSTSUBSCRIPT italic_X end_POSTSUBSCRIPT start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTof the variational posteriorqϕ(𝐳∣𝐱)subscript𝑞italic-ϕconditional𝐳𝐱q_{\phi}(\mathbf{z}\mid\mathbf{x})italic_q start_POSTSUBSCRIPT italic_ϕ end_POSTSUBSCRIPT ( bold_z ∣ bold_x ). The decoder consists of a fully connected layer followed by an LSTM layer. | 2501.14615v2 | 2 |
|
Figure 3:Fluorescence trace reconstruction.A.Three representative examples of the noisy trace𝐱𝐱\mathbf{x}bold_x(blue) withσt=1subscript𝜎𝑡1\sigma_{t}=1italic_σ start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT = 1and reconstructed𝐱^^𝐱\widehat{\mathbf{x}}over^ start_ARG bold_x end_ARG(red) for an AVAE withD1=128subscript𝐷1128D_{1}=128italic_D start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 128,D2=8subscript𝐷28D_{2}=8italic_D start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT = 8, andβKL=0.5subscript𝛽KL0.5\beta_{\text{KL}}=0.5italic_β start_POSTSUBSCRIPT KL end_POSTSUBSCRIPT = 0.5.B.Mean-squared error (MSE) (×103absentsuperscript103\times 10^{3}× 10 start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT) of the reconstruction for different levels of noiseσtsubscript𝜎𝑡\sigma_{t}italic_σ start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTand latent dimensionD2subscript𝐷2D_{2}italic_D start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT(color indicates the corresponding model). | 2501.14615v2 | 3 |
|
Figure 4:t-SNE of learned representation colored according to neuron firing type (A), connectivity (B), parameterv0subscript𝑣0v_{0}italic_v start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT(C), and parameterΔuΔ𝑢\Delta uroman_Δ italic_u(D). Neurons are grouped according to connectivity and firing type, largely driven by parametersv0subscript𝑣0v_{0}italic_v start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTandΔuΔ𝑢\Delta uroman_Δ italic_u. | 2501.14615v2 | 4 |
|
Figure 5:Quality of latent representation in simulated data.A.Adjusted Rand index (ARI) of the DBSCAN clustering under different levels of noiseσtsubscript𝜎𝑡\sigma_{t}italic_σ start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTand latent dimensionD2subscript𝐷2D_{2}italic_D start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT(color indicates the corresponding model).B.Correlation between the parameters of the dynamical model and the latent representation. Parametersa𝑎aitalic_aandb𝑏bitalic_bdo not correlate with any coordinate, whereas the informative parametersv0subscript𝑣0v_{0}italic_v start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTandΔuΔ𝑢\Delta uroman_Δ italic_ucorrelate in an antagonistic way with the first seven coordinates of𝐳𝐳\mathbf{z}bold_z. | 2501.14615v2 | 5 |
|
Figure 6:Reconstruction performance of the AVAE model (D1=128subscript𝐷1128D_{1}=128italic_D start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 128,D2=8subscript𝐷28D_{2}=8italic_D start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT = 8) for thein vitrotauopathy model data.A.Representative examples of originalX𝑋Xitalic_X(blue) and reconstructedX^^𝑋\widehat{X}over^ start_ARG italic_X end_ARG(red) signals in control (top) and diseased (bottom) neurons from the validation subset.B.t-SNE embedding of calcium recordings from anin vitrotauopathy model, showing control neurons (green) and diseased neurons (orange) from the validation subset. The latter exhibits a higher degree of heterogeneity than the former, probably due to connectivity differences. | 2501.14615v2 | 6 |
|
Figure 1:(A) Schematic depiction of data generation. CI-HRIR = Cochlear implants Head Impulse Response; RIR = Room impulse response; BRIR = Binaural Room Impulse Response. (B) An acoustic scene depicting talkers positioned around a listener located in a room. (C) Schematic example of an acoustic scene with two talkers and a listener wearing bilateral CIs. Each CI has three microphones: F = front microphone, B = back microphone and T = T-microphone. | 2501.14610v1 | 1 |
|
Figure 2:(A) Reverberation time (T60subscript𝑇60T_{60}italic_T start_POSTSUBSCRIPT 60 end_POSTSUBSCRIPT) as a function of room size. Each circle represents a single room. (B) Effect of reverberation and spatialization on speech mixtures. Top panes show waveforms of a dry, non-spatial speech mixture and a spatial, reverberant version of the same speech mixture. For illustration, bottom panes depict spectrograms (but note that models are trained directly on the waveform). (C) Presence of implicit spatial cues in real-world acoustic scenes. Top pane shows an example of a two-talker mixture in a spatial, reverberant scene. One talker is at -90∘, the other talker is at +90∘. Bottom pane visualizes the corresponding interaural level differences for this speech mixture. | 2501.14610v1 | 2 |
|
Figure 3:(A) SuDoRM-RF[24]implementation in present study. Top row depicts SuDoRM-RF approach for input configurations consisting only of the speech mixture (i.e. waveform). Bottom row depicts our SuDoRM-RF approach for input configurations consisting of speech mixture (i.e. waveform) and IPD as an auxiliary feature. (B) Schematic overview of all input configurations. Color indicates the presence of implicit (blue) and combined implicit and explicit (green) spatial cues. A higher saturation signifies stronger spatial cues. | 2501.14610v1 | 3 |
|
Figure 4:Speech separation performance as a function of spatial separation angle between talkers. Each panel depicts the results for one input configuration (Fig. 3B, TableI). Colors indicate implicit (blue) and combined implicit and explicit (green) spatial information load, with higher saturation indicating higher spatial information load (similar to Fig. 3B). Bars represent average SI-SDR, error bars reflect standard error of the mean. As a reference, the gray dashed line indicates the average SI-SDR at a separation angle of 0°, that is, no spatial distance. Asterisks reflect a statistically significant difference in SI-SDR at that relevant separation angle and the 0° reference: * =p<<<0.05, ** =p<<<0.01. | 2501.14610v1 | 4 |
|
Figure 5:SI-SDR as a function of talker gender pairing and spatial distance for various input configurations. Bars represent average SI-SDR at small spatial distances (0° and 15°, filled bars) or and large spatial distances (60° and 90°, open bars) across the different gender pairings (red = two female talkers; blue = two male talkers; purple = one male and one female talker). Asterisks reflect a statistically significant difference (Kruskal-Wallis H tests, FDR corrected): *** =p<<<0.001; ** =p<<<0.01. | 2501.14610v1 | 5 |
|
Figure 1:Initially, agents are non-collaborative (I), resulting in degraded performance. Collaborative Feature Alignment (CFA) enables collaboration among heterogeneous agents through a two-step process (II): training a protocol network and training local adapters and reverters. Theprotocol networkfacilitates communication betweenAgent 1,Agent 2, andAgent 3, each with heterogeneous models and features. Gradient-colored feature maps represent features adapted or reverted between domains. After CFA implementation, agents become collaborative (III) with improved performance. | 2501.18616v1 | 1 |
|
Figure 2:Training efficiency comparison of our framework and existing heterogeneous CP frameworks across a number of heterogeneous agents. | 2501.18616v1 | 2 |
|
Figure 3:Ablation studies on the OPV2V dataset: (a) Model performance across different BEV feature channel sizes. (b) Performance comparison of various adapter and reverter architectures. (c) Performance results using different combinations of loss function components (Lfsubscript𝐿𝑓L_{f}italic_L start_POSTSUBSCRIPT italic_f end_POSTSUBSCRIPTandLdsubscript𝐿𝑑L_{d}italic_L start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT). | 2501.18616v1 | 3 |
|
Figure 5:Visualization of feature maps and model outputs before and after Collaborative Feature Alignment (CFA) for two scenes with different agents and tasks. For 3D object detection,greenboxes indicate the ground truth labels andredboxes indicate the predictions. CFA enhances feature clarity and information preservation, resulting in improved perception accuracy across heterogeneous agents. | 2501.18616v1 | 5 |
|
Figure A2:Architectural comparison of collaborative perception frameworks: existing approaches versus our proposed STAMP method.Blue boxesrepresent models with fixed parameters, whilered boxesindicate models whose parameters are trained during the collaboration process. | 2501.18616v1 | 7 |
|
Figure A3:Comparison of collaborative perception systems: (Left) Single-group system where all agents collaborate within one group. (Middle) Multi-group single-model system allowing agents to join only one of multiple collaboration groups. (Right) Multi-group multi-model system enabling agents to participate in multiple collaboration groups simultaneously. The figure illustrates how different system architectures impact agent interactions and group formations in autonomous driving scenarios. | 2501.18616v1 | 8 |
|
Figure A4:Visualization of feature maps and inference results before and after Collaborative Feature Alignment (CFA) in a three-agent scene.Ai→Aj→subscript𝐴𝑖subscript𝐴𝑗A_{i}\rightarrow A_{j}italic_A start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT → italic_A start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPTdenotes the feature map aligned from agenti𝑖iitalic_i’s domain to agentj𝑗jitalic_j’s domain, also represented asFijsubscript𝐹𝑖𝑗F_{ij}italic_F start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPT. | 2501.18616v1 | 9 |
|
Figure A5:Visualization of feature maps and inference results before and after Collaborative Feature Alignment (CFA) in a four-agent scene.Ai→Aj→subscript𝐴𝑖subscript𝐴𝑗A_{i}\rightarrow A_{j}italic_A start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT → italic_A start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPTdenotes the feature map aligned from agenti𝑖iitalic_i’s domain to agentj𝑗jitalic_j’s domain, also represented asFijsubscript𝐹𝑖𝑗F_{ij}italic_F start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPT. | 2501.18616v1 | 10 |
|
Figure 1:(a) The SOTA method[22]fails to handle the descriptions involving composite object attributes and spatial locations.
(b) GroundingDINO[21]can neither understand the temporal dynamics nor perform pixel-wise segmentation. | 2501.14607v1 | 1 |
|
Figure 2:Overall architecture of ReferDINO. Modules colored inblueare borrowed from GroundingDINO, while those inredare newly introduced in this work.
Based on the frame-wise object features{𝒪t}t=1Tsubscriptsuperscriptsuperscript𝒪𝑡𝑇𝑡1\{\mathcal{O}^{t}\}^{T}_{t=1}{ caligraphic_O start_POSTSUPERSCRIPT italic_t end_POSTSUPERSCRIPT } start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t = 1 end_POSTSUBSCRIPT, ourobject-consistent temporal enhancerleverages the cross-modal text features to enable inter-frame object interaction.
Then, ourgrounding-guided deformable mask decoderproduces masks for the candidate objects, conditioned on location predictions, cross-modal text features and high-resolution feature maps.
To further improve the efficiency in video processing, we introduce aconfidence-aware query pruningstrategy in the cross-modal decoder.
Best view in color. | 2501.14607v1 | 2 |
|
Figure 3:Illustration of ourobject-consistent temporal enhancer, wherefclstsubscriptsuperscript𝑓𝑡𝑐𝑙𝑠f^{t}_{cls}italic_f start_POSTSUPERSCRIPT italic_t end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c italic_l italic_s end_POSTSUBSCRIPTis the cross-modal sentence feature oft𝑡titalic_t-th frame. | 2501.14607v1 | 3 |
|
Figure 4:Illustration of ourconfidence-aware query pruning, whereqisubscript𝑞𝑖q_{i}italic_q start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTandtisubscript𝑡𝑖t_{i}italic_t start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTindicatei𝑖iitalic_i-th query and text token, respectively. | 2501.14607v1 | 4 |
|
Figure 7:Qualitative impacts ofα𝛼\alphaitalic_αin memory-augmented tracker. We use ✕ to highlight the incorrect results. | 2501.14607v1 | 7 |
|
Figure 1:Illustration of our approach using pseudo-dense points for domain generalization of LiDAR semantic Segmentation in autonomous driving (the blue sphere represents the position of the ego vehicle). | 2501.14605v1 | 1 |
|
Figure 5:Illustration of the trail phenomenon. On the left, a section of a point cloud made up of 5 consecutive scans; on the right, a section composed of 20 consecutive scans. | 2501.14605v1 | 5 |
|
Figure 1:Comparison between inverse evolution and forward evolution on heat diffusion equation withΔt=0.05Δ𝑡0.05\Delta t=0.05roman_Δ italic_t = 0.05. Inverse evolution produces more accurate data pairs (U0,U1subscript𝑈0subscript𝑈1U_{0},U_{1}italic_U start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT , italic_U start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT) as the order increases, while forward evolution yields inaccurate data pairs (U1,U2subscript𝑈1subscript𝑈2U_{1},U_{2}italic_U start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , italic_U start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT) across all schemes. | 2501.14604v1 | 1 |
|
Figure 2:Inverse evolution between original data and preprocessed data. The first row shows the inverse evolution of the original data, while the second row presents the results based on preprocessed data. | 2501.14604v1 | 2 |
|
Figure 3:Generated data for Allen-Cahn equation withϵ=0.01italic-ϵ0.01\epsilon=0.01italic_ϵ = 0.01andΔt=0.5Δ𝑡0.5\Delta t=0.5roman_Δ italic_t = 0.5. The first two columns are generated inputs and outputs from the inverse evolution, respectively. The last column shows the true solutions of the generated input. | 2501.14604v1 | 3 |
|
Figure 4:Generated data for Navier-Stokes equation withν=0.001𝜈0.001\nu=0.001italic_ν = 0.001andΔt=0.5Δ𝑡0.5\Delta t=0.5roman_Δ italic_t = 0.5. The first two columns are generated inputs and outputs from the inverse evolution, respectively. The last column shows the true solutions of the generated input. | 2501.14604v1 | 4 |
|
Figure 5:Predictions of FNO trained on different datasets for Allen-Cahn equation (ϵ=0.05italic-ϵ0.05\epsilon=0.05italic_ϵ = 0.05). (a) inputs with noise (b) 1000 data pairs (c) 10000 data pairs (d) 1000 data pairs + 1000 generated data pairs (e) 5000 data pairs + 5000 generated data pairs (f) ground truth | 2501.14604v1 | 5 |
|
Figure 1:The considered system model operating within a grid environment, consists of a UAV that serves multiple, fixed located IoT devices with limited power. | 2501.14603v1 | 1 |
|
Figure 4:Meta-testing convergence compared to random weights initialization with an environment withλ=300𝜆300\lambda=300italic_λ = 300. | 2501.14603v1 | 7 |
|
Figure 6:Testing the proposed meta-RL algorithm over11111111different environments (with differentλ𝜆\lambdaitalic_λvalues) regarding the achieved AoI and transmission power. | 2501.14603v1 | 9 |
|
Fig. 1:Performance analysis of various distance metric𝚍psubscript𝚍𝑝{\mathtt{d}}_{p}typewriter_d start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT. | 2501.14593v1 | 1 |
|
Fig. 1:Symmetric Rotation-Equivariant (SRE) U-Net.Compared to standard U-Net, SRE U-Net utilizes a parameter-efficient symmetric kernel to enable rotational equivariance. Difference image maps between the output of the original and rotated (90°) images illustrate SRE U-Net’s equivariant property by maintaining consistent feature response after rotation on a 2D fundus image. | 2501.14592v1 | 1 |
|
Fig. 2:Performance Across Rotation Degree.We plot Dice and AUC across each rotation degree to visualize the influence of small rotation on the model’s performance. | 2501.14592v1 | 2 |
|
Fig. 3:Qualitative Results.An example fundus image is rotated and fed into the models and the outputs are rotated back to compare with the output from the original image. Difference maps visualize the difference in the predictions under different rotational conditions and we quantitatively summarize this error using Mean Squared Error (MSE). Our proposed method shows consistent results due to its rotational equivariance property. | 2501.14592v1 | 3 |
|
Figure 1:More noise may not increaseDist(ℋ,ℋ′)Distℋsuperscriptℋ′\operatorname{\text{Dist}}(\mathcal{H},\mathcal{H}^{\prime})dist ( caligraphic_H , caligraphic_H start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ). | 2501.10606v1 | 1 |
|
Figure 1:We aim to leverage Multimodal Large Language Models (MLLMs) to assist traffic accident analysis. Domain knowledge and traffic camera footage construct the task-specific prompt. The MLLMs generate structured responses. | 2501.10604v1 | 1 |
|
Figure 2:The overall architecture of our SeeUnsafe framework. It processes video inputs by generating visual prompts for N frames and splitting them into K clips. Textual prompts guide the MLLM agent in generating structured responses for each clip. Outputs are then aggregated using a severity-based protocol to deliver overall event classification, semantic descriptions, and object grounding in critical events. | 2501.10604v1 | 2 |
|
Figure 3:Workflow for generating visual prompts. An open-set object detection model identifies the bounding boxes of objects of interest and assigns a unique ID to each object in the given frame. These bounding boxes are then used to guide segmentation models, which generate instance masks and track objects across frames. Only the object contours are overlaid, while the complete masks are stored for future indexing. | 2501.10604v1 | 3 |
|
Figure 4:We show two types of potential failures in structured responses for theobject_descriptioncomponent when describing road users involved in a collision. Correct contents are highlighted ingreen, while incorrect ones are inred. (Best viewed in color and see also Table1). | 2501.10604v1 | 4 |
|
Figure 5:Word cloud of the 100 most frequent words in the processed WTS dataset captions. Larger words appear more frequently. | 2501.10604v1 | 5 |
|
Figure 7:Visualization of semantic evaluation results of correctly classified videos with and without visual prompt based on GPT-4o. The x-axis is thescene_contextscore, the y-axis is theobject_descriptionscore, and thejustificationscore defines the color. There are three marker types:∗*∗for collision,×\times×for near-miss, and+++for normal videos, respectively. The bigger markers in black are the centroid of the corresponding video classes. We provide detailed quantitative results in Table3. | 2501.10604v1 | 7 |
|
Figure 8:Confusion matrix for accident-aware video classification using different MLLM agents within the SeeUnsafe framework with and without visual prompts. | 2501.10604v1 | 8 |
|
Figure 9:Evaluation of generated responses in both correct and incorrect classifications. We use S for scene context, O for object description, and J for justification. | 2501.10604v1 | 9 |
|
Figure 10:Qualitative evaluation of BLEU, ROUGE-L, and IMS for responses of SeeUsafe with a GPT-4o mini agent. | 2501.10604v1 | 10 |
|
Figure 11:Qualitative evaluation of BLEU, ROUGE-L, and IMS for responses of SeeUsafe with anLLaVA-NeXTinterleavesubscriptLLaVA-NeXT𝑖𝑛𝑡𝑒𝑟𝑙𝑒𝑎𝑣𝑒\text{LLaVA-NeXT}_{interleave}LLaVA-NeXT start_POSTSUBSCRIPT italic_i italic_n italic_t italic_e italic_r italic_l italic_e italic_a italic_v italic_e end_POSTSUBSCRIPTagent. | 2501.10604v1 | 11 |
|
Figure 12:Qualitative evaluation of BLEU, ROUGE-L, and IMS for responses of SeeUsafe with anLLaVA-NeXTvideosubscriptLLaVA-NeXT𝑣𝑖𝑑𝑒𝑜\text{LLaVA-NeXT}_{video}LLaVA-NeXT start_POSTSUBSCRIPT italic_v italic_i italic_d italic_e italic_o end_POSTSUBSCRIPTagent. | 2501.10604v1 | 12 |
|
Figure 13:Qualitative comparison of IMS computed by GPT-4o and GPT-4 for responses of SeeUsafe with a GPT-4o agent. | 2501.10604v1 | 13 |
|
Figure 1:Location of the 3,060 LiDAR data points in the Amazon domain used in the validation of the Amazon forest tree canopy height model (shown in gold). As at least one image per flight was used in validation, the validation locations approximately represent the locations of all the LiDAR flights in the Amazon used in the study. The Amazon domain is partitioned into four regions according toFeldpausch et al. (2011). | 2501.10600v1 | 1 |
|
Figure 2:U-Net model architecture used for canopy height estimation from Planet NICFI images, adapted fromRonneberger et al. (2015). The number of channels is indicated above the cuboids, and the vertical numbers indicate the row and column sizes in pixels. The operations (convolutions, skip connections, max pooling, and upsampling) performed in each layer and their sizes are indicated by the colored arrows. | 2501.10600v1 | 2 |
|
Figure 3:Comparison of predicted versus observed height (m) for the 3,436 validation areas represented as density scatterplots for our canopy height model (a), Tolan’s model (b), and Lang’s model (c). Tolan’s model and Lang’s model, with native spatial resolutions of 0.5 m and 10 m respectively, were warped to our 4.78 m spatial resolution using the median and nearest neighbor algorithms, respectively. Each plot contains approximately 63 million points. The 1:1 line and the mean absolute error are depicted in white. | 2501.10600v1 | 3 |
|
Figure 4:Distribution of the observed height in the validation sample (a); distribution of the predicted height in the validation sample for our model, Tolan’s model, and Lang’s model (b); and distribution of the differences in predicted height in the validation sample between our model and Tolan’s model, and between our model and Lang’s model (c), respectively. | 2501.10600v1 | 4 |
|
Figure 5:Example of canopy height models observed in the validation dataset (column 1), predicted from our canopy height model (column 2), predicted from Tolan’s model (column 3), and predicted from Lang’s model (column 4). Tolan’s model, Pauls’s model and Lang’s model, whose native spatial resolutions are 0.5 m, 10 m and 10 m, respectively, have been warped to our 4.78 m spatial resolution using the median and nearest neighbor algorithms, respectively. | 2501.10600v1 | 5 |
|
Figure 6:Example of canopy height models observed in the validation dataset (column 1), predicted from our canopy height model (column 2), predicted from Tolan’s model (column 3), and predicted from Lang’s model (column 4). Tolan’s model, Pauls’s model and Lang’s model, whose native spatial resolutions are 0.5 m, 10 m and 10 m, respectively, have been warped to our 4.78 m spatial resolution using the median and nearest neighbor algorithms, respectively. | 2501.10600v1 | 6 |
Subsets and Splits