SciForma: Structure-Faithful Generation of Scientific Diagrams
Paper • 2607.18091 • Published • 24
Datasets:
split stringclasses 1
value | index int32 0 599 | prompt large_stringlengths 1.75k 4.39k | width int32 576 2.05k | height int32 512 1.73k | arxiv_id stringlengths 10 10 | year stringdate 2015-01-01 00:00:00 2025-01-01 00:00:00 | image_path stringlengths 38 49 | gt_image imagewidth (px) 576 2.05k | rubric_components large_stringlengths 652 3.68k | rubric_arrows large_stringlengths 2 2.96k | rubric_text large_stringlengths 2 3.37k |
|---|---|---|---|---|---|---|---|---|---|---|---|
hard | 0 | The figure illustrates a human parsing pipeline composed of four main stages: Segment Generation, Segment Feature Extraction, Segment Ranking and Selection, and Part Assembling by And-or Graph. The global layout is left-to-right, with data flowing from an input image of a person on the far left through successive proce... | 1,472 | 704 | 1508.03881 | 2015 | 1508.03881/1508.03881_1_img_SINGLE.png | [{"name": "Input person image", "description": "Original image of a walking person at the far left, serving as the pipeline input."}, {"name": "Pose estimation image", "description": "Annotated person image with body pose/skeleton cues used to support later feature extraction and selection."}, {"name": "Segment Generat... | [{"source": "Input person image", "target": "Segment Generation", "label": "generate candidate segments"}, {"source": "Input person image", "target": "Pose estimation image", "label": "derive pose cues"}, {"source": "Pose estimation image", "target": "Segment Generation", "label": "support segment generation"}, {"sourc... | [{"text": "Segment Generation", "location": "top-left stage box"}, {"text": "Pose-based Segment Feature", "location": "lower middle auxiliary box"}, {"text": "Segment Ranking and Selection", "location": "upper middle-right stage box"}, {"text": "Part Assembling by And-or Graph", "location": "top-right large box"}, {"te... | |
hard | 1 | The figure presents two side-by-side diagrams, labeled (a) and (b), illustrating the application of a proposed method in convolutional neural networks (CNNs) using stacked 'quasi-hexagonal' kernels. Both diagrams depict a three-layered CNN structure, with layers labeled as Layer l, Layer l+1, and Layer l+2, arranged ve... | 1,600 | 640 | 1511.09231 | 2015 | 1511.09231/1511.09231_3_MULTIFIGURE_COMPOSED.png | [{"name": "Layer l", "description": "Bottom grid layer representing the input feature map in both subfigures.", "importance": "primary"}, {"name": "Kernel Pattern 'U'", "description": "A small quasi-hexagonal convolution kernel with a U-shaped footprint (colored blue in the middle grid of (a) and orange in the bottom g... | [] | [{"text": "(b)", "location": "top-left corner of right diagram", "importance": "secondary"}, {"text": "Layer l+2", "location": "above the top grid in each subfigure", "importance": "secondary"}, {"text": "neuron N(i,j,c)", "location": "arrow pointing to red square in Layer l+2", "importance": "secondary"}, {"text": "co... | |
hard | 2 | The figure presents four distinct methodologies (labeled a, b, c, d) for training deep convolutional neural networks (CNNs) to classify 2D radiology images into different topic categories, using text from radiology reports as supervisory signals. Each panel illustrates a different level of text granularity and topic mo... | 1,088 | 960 | 1505.00670 | 2015 | 1505.00670/1505.00670_5_img_SINGLE.png | [{"name": "Documents (radiology reports)", "description": "Corpus of radiology reports that serve as raw text input.", "importance": "primary"}, {"name": "NLP preprocessing", "description": "Natural-language pipeline that extracts image references and relevant sentences from the reports.", "importance": "primary"}, {"n... | [{"source": "Documents (radiology reports)", "target": "NLP preprocessing", "label": null, "importance": "primary"}, {"source": "Documents (radiology reports)", "target": "LDA document topic clustering", "label": null, "importance": "primary"}, {"source": "NLP preprocessing", "target": "Sentences mentioning the finding... | [{"text": "(a)", "location": "bottom left quadrant", "importance": "secondary"}, {"text": "(b)", "location": "bottom right quadrant", "importance": "secondary"}, {"text": "(c)", "location": "bottom left quadrant", "importance": "secondary"}, {"text": "(d)", "location": "bottom right quadrant", "importance": "secondary"... | |
hard | 3 | The figure presents a complete flow diagram of the Recursive Context Propagation Network (RCPN) for semantic segmentation, structured into two main stages: Visual Feature Extraction and Recursive Context Propagation Network (RCPN), separated by a dashed vertical boundary.
[1] Global Layout and Structure:
The diagram i... | 1,728 | 576 | 1503.02725 | 2015 | 1503.02725/1503.02725_0_img_SINGLE.png | [{"name": "Input Image", "description": "Raw scene image fed into the network (labeled I).", "importance": "primary"}, {"name": "F_CNN", "description": "Convolutional feature–extraction block producing a dense feature map.", "importance": "primary"}, {"name": "Feature Map V", "description": "Intermediate visual-feature... | [{"source": "Input Image", "target": "F_CNN", "label": null, "importance": "primary"}, {"source": "F_CNN", "target": "Feature Map V", "label": null, "importance": "primary"}, {"source": "Feature Map V", "target": "Feature Fusion (+)", "label": null, "importance": "secondary"}, {"source": "Superpixels", "target": "Featu... | [{"text": "Visual Feature Extraction", "location": "Caption under left dashed box", "importance": "secondary"}, {"text": "Recursive Context Propagation Network (RCPN)", "location": "Caption under right dashed box", "importance": "secondary"}, {"text": "Bypass error", "location": "Red dashed arrow from F_sem to F_dec", ... | |
hard | 4 | The figure presents an overview of the Horde architecture, a framework for learning an ensemble of reward shapings in reinforcement learning. The global layout is structured as a flowchart with distinct modules arranged from top-left to bottom-right, depicting a data and control flow from the environment through shapin... | 1,024 | 1,024 | 1502.03248 | 2015 | 1502.03248/1502.03248_0_img_SINGLE.png | [{"name": "Environment", "description": "Oval representing the external system that provides states s and rewards R and receives actions a and a'.", "importance": "primary"}, {"name": "Shaping", "description": "Block that combines the environment signals (s, R) with designer-provided potential Φ to produce shaped poten... | [{"source": "Environment", "target": "Shaping", "label": "R", "importance": "primary"}, {"source": "Environment", "target": "Shaping", "label": "s", "importance": "primary"}, {"source": "Designer", "target": "Shaping", "label": "Φ", "importance": "secondary"}, {"source": "Shaping", "target": "Scaling", "label": "Φ₁…Φℓ"... | [{"text": "Behavior", "location": "text on green oval surrounding Environment", "importance": "secondary"}, {"text": "s", "location": "arrow from Behavior to Environment inside oval", "importance": "secondary"}, {"text": "a", "location": "arrow from Environment back to Behavior inside oval", "importance": "secondary"},... | |
hard | 5 | The figure illustrates a deep learning framework for fine-grained classification using triplet-based training with shared parameters across three convolutional neural networks (CNNs). The global layout is left-to-right, starting with the input data on the far left, progressing through three parallel CNN branches, follo... | 1,920 | 512 | 1512.02895 | 2015 | 1512.02895/1512.02895_1_img_SINGLE.png | [{"name": "Training Data", "description": "Stack of color-coded ref/pos/neg samples that serves as network input.", "importance": "primary"}, {"name": "Parameter-Sharing CNN", "description": "Three parallel CNN branches (R, P, N) that share weights and process different inputs.", "importance": "primary"}, {"name": "ℓ₂ ... | [{"source": "Training Data", "target": "Parameter-Sharing CNN", "label": null, "importance": "primary"}, {"source": "Parameter-Sharing CNN", "target": "ℓ₂ Normalization", "label": null, "importance": "primary"}, {"source": "ℓ₂ Normalization", "target": "Structured Label Module", "label": null, "importance": "primary"},... | [{"text": "ref pos neg", "location": "legend above Training Data stack", "importance": "secondary"}, {"text": "R", "location": "inside blue CNN branch", "importance": "secondary"}, {"text": "P", "location": "inside green CNN branch", "importance": "secondary"}, {"text": "N", "location": "inside red CNN branch", "impo... | |
hard | 6 | The figure illustrates the initial synaptic weight configuration for a neural network controlling a robot's movement behavior, specifically designed to enforce an attraction response. The global layout is a two-layer feedforward network with four input neurons at the top and two output motor neurons at the bottom. The ... | 1,792 | 576 | 1502.06096 | 2015 | 1502.06096/1502.06096_40_img_SINGLE.png | [{"name": "Left Food Sensor Neurons", "description": "Neuron group sensing food on the organism’s left side.", "importance": "primary"}, {"name": "Right Food Sensor Neurons", "description": "Neuron group sensing food on the organism’s right side.", "importance": "primary"}, {"name": "Left Container Neurons", "descripti... | [{"source": "Left Food Sensor Neurons", "target": "Left Motor Neurons", "label": "Synapses Set to Maximum", "importance": "primary"}, {"source": "Left Food Sensor Neurons", "target": "Right Motor Neurons", "label": "Synapses Set to Zero", "importance": "primary"}, {"source": "Right Food Sensor Neurons", "target": "Righ... | [{"text": "Synapses Set to Maximum", "location": "Inside legend box beside solid black arrow", "importance": "secondary"}, {"text": "Synapses Set to Zero", "location": "Inside legend box beside dotted blue arrow", "importance": "secondary"}] | |
hard | 7 | The figure presents an end-to-end neural network architecture for image captioning, structured into two main horizontal pathways: a top-level processing flow and a detailed Localization Layer at the bottom. The global layout begins on the left with an input image of dimensions 3×W×H, depicted as a photograph of two cat... | 1,472 | 704 | 1511.07571 | 2015 | 1511.07571/1511.07571_1_img_SINGLE.png | [{"name": "Input Image", "description": "Original RGB image input (3 × W × H).", "importance": "primary"}, {"name": "CNN", "description": "Convolutional neural network extracting low-level convolutional features.", "importance": "primary"}, {"name": "Conv features", "description": "Feature map tensor produced by the CN... | [{"source": "Input Image", "target": "CNN", "label": null, "importance": "primary"}, {"source": "CNN", "target": "Conv features", "label": null, "importance": "primary"}, {"source": "Conv features", "target": "Localization Layer", "label": null, "importance": "primary"}, {"source": "Localization Layer", "target": "Regi... | [{"text": "Image: 3 × W × H", "location": "above the input image", "importance": "primary"}, {"text": "Conv features: C × W′ × H′", "location": "above the light-blue cube", "importance": "secondary"}, {"text": "Region features: B × C × X × Y", "location": "above the vertical blue bar near recognition network", "importa... | |
hard | 8 | The figure presents an overview of a joint image-text topic detection and tracking framework for news content. The global layout is structured as a top-down pipeline divided into three main stages: preprocessing, topic detection, and topic tracking, separated by dashed horizontal lines. At the top, two input sources ar... | 960 | 1,088 | 1512.04701 | 2015 | 1512.04701/1512.04701_0_img_SINGLE.png | [{"name": "News Videos", "description": "Input box that contains a montage of video frames from news broadcasts.", "importance": "primary"}, {"name": "News Captions", "description": "Input box that shows transcribed caption text accompanying the news videos.", "importance": "primary"}, {"name": "Joint Image-Text Topic ... | [] | [{"text": "Topics Detected in Time Period 1", "location": "above the left And-Or tree inside the Topic Detection block", "importance": "secondary"}, {"text": "Topic 1-1", "location": "root of the left And-Or tree", "importance": "secondary"}, {"text": "Text", "location": "left child of Topic 1-1 and Topic M-1", "import... | |
hard | 9 | The figure illustrates an end-to-end deep neural network architecture designed for generating natural language descriptions from video sequences. The global layout is horizontal and sequential, depicting a pipeline that processes multiple video frames over time, each frame being analyzed at multiple spatial scales befo... | 1,728 | 576 | 1505.05914 | 2015 | 1505.05914/1505.05914_1_img_SINGLE.png | [{"name": "Input Image (Scale 1)", "description": "First resolution of the video frame showing the scene.", "importance": "primary"}, {"name": "Input Image (Scale 2)", "description": "Second (medium) resolution of the same frame.", "importance": "primary"}, {"name": "Input Image (Scale 3)", "description": "Third (high)... | [{"source": "Input Image (Scale 1)", "target": "conv1-5 (scale 1)", "label": null, "importance": "primary"}, {"source": "conv1-5 (scale 1)", "target": "fc6 (scale 1)", "label": "1×K", "importance": "secondary"}, {"source": "fc6 (scale 1)", "target": "fc7 (scale 1)", "label": "1×K", "importance": "secondary"}, {"source"... | [{"text": "scale 1", "location": "left of first image row", "importance": "primary"}, {"text": "scale 2", "location": "left of second image row", "importance": "primary"}, {"text": "scale 3", "location": "left of third image row", "importance": "primary"}, {"text": "...", "location": "beneath third image, indicating mo... | |
hard | 10 | The figure presents two architectural diagrams, labeled (a) tagCNN and (b) inCNN, illustrating variations of a convolutional neural network (CNN)-based encoder for a joint language model (LM) task, likely for named entity recognition or relation extraction. Both diagrams share a common top-level structure: a deep neura... | 1,344 | 768 | 1503.01838 | 2015 | 1503.01838/1503.01838_0_MULTIFIGURE_COMPOSED.png | [{"name": "tagCNN", "description": "Left–hand model variant illustrating the basic convolution-based encoder.", "importance": "primary"}, {"name": "inCNN", "description": "Right–hand model variant that augments the encoder with the additional context vector h(e2,e3,e4).", "importance": "primary"}, {"name": "Input Token... | [{"source": "Input Token Sequence (z1–z6)", "target": "Convolution & Gating (tagCNN)", "label": null, "importance": "primary"}, {"source": "Convolution & Gating (tagCNN)", "target": "CNN Output Embeddings (tagCNN)", "label": null, "importance": "primary"}, {"source": "CNN Output Embeddings (tagCNN)", "target": "DNN (ta... | [{"text": "convolution & gating", "location": "inside both convolution blocks", "importance": "secondary"}, {"text": "\"presidential\"", "location": "above green output box", "importance": "secondary"}, {"text": "智利 举行 国会 与 总统 选举", "location": "caption under z1–z6 tokens", "importance": "secondary"}, {"text": "z1–z6", ... | |
hard | 11 | The figure illustrates a hybrid preference prediction scheme that integrates visual content analysis with collaborative filtering, enabling predictions even when no past user ratings are available. The global layout is horizontal, progressing from left to right through five main stages: Query, Visual space, Latent spac... | 1,408 | 704 | 1510.07867 | 2015 | 1510.07867/1510.07867_1_img_SINGLE.png | [{"name": "Visual Query Input", "description": "An image supplied as the query example.", "importance": "primary"}, {"name": "Past Ratings Input", "description": "Historical thumbs-up / thumbs-down ratings supplied by the user.", "importance": "primary"}, {"name": "Visual Space", "description": "Grid of candidate image... | [{"source": "Visual Query Input", "target": "Visual Space", "label": "Nearest neighbor", "importance": "primary"}, {"source": "Visual Space", "target": "Latent Space (Q)", "label": "Regression", "importance": "primary"}, {"source": "Past Ratings Input", "target": "Latent Space (Q)", "label": "Matrix factorization", "im... | [{"text": "Query", "location": "above the two input options", "importance": "secondary"}, {"text": "Visual", "location": "vertical text beside image query", "importance": "secondary"}, {"text": "OR", "location": "between the two input options", "importance": "secondary"}, {"text": "Past ratings", "location": "vertical ... | |
hard | 12 | The figure illustrates the architecture of a system combining Redis with KVolve, designed to manage key-value operations with versioning and update specifications. The global layout is divided into two main regions: on the left, three client entities (ClientX, ClientU, ClientY) are shown initiating operations; on the r... | 1,088 | 960 | 1506.08800 | 2015 | 1506.08800/1506.08800_0_img_SINGLE.png | [{"name": "ClientX", "description": "Top-left client issuing the operation “get k:x” and holding local state {k:v0}.", "importance": "primary"}, {"name": "ClientU", "description": "Middle client that installs a schema/data update with the command “upd k: v0 → v1”.", "importance": "primary"}, {"name": "ClientY", "descri... | [{"source": "ClientX", "target": "Redis + KVolve", "label": "get k:x", "importance": "primary"}, {"source": "ClientX", "target": "Redis + KVolve", "label": "close", "importance": "secondary"}, {"source": "ClientU", "target": "upd spec", "label": "install", "importance": "primary"}, {"source": "ClientY", "target": "Redi... | [{"text": "v0 → v1", "location": "inside upd k: box in ClientU", "importance": "secondary"}] | |
hard | 13 | The figure illustrates a two-stage deep learning framework for saliency detection, structured into two main phases: coarse representation learning and fine representation learning. The global layout is left-to-right, beginning with an input image on the far left and progressing through two sequential processing stages ... | 1,728 | 576 | 1511.04192 | 2015 | 1511.04192/1511.04192_1_img_SINGLE.png | [{"name": "Input Image", "description": "Original color photograph fed into the pipeline.", "importance": "primary"}, {"name": "Superpixel Segmentation", "description": "Image over-segmented into superpixels; shown beneath the input photo.", "importance": "secondary"}, {"name": "Coarse Representation Learning Block", "... | [{"source": "Input Image", "target": "Coarse Representation Learning Block", "label": null, "importance": "primary"}, {"source": "Superpixel Segmentation", "target": "Coarse Representation Learning Block", "label": null, "importance": "secondary"}, {"source": "Coarse Representation Learning Block", "target": "Coarse Fe... | [{"text": "Coarse representation learning", "location": "Top of blue dashed box", "importance": "primary"}, {"text": "Fine representation learning", "location": "Top of green dashed box", "importance": "primary"}, {"text": "ISS & ISV", "location": "Between coarse and fine modules", "importance": "primary"}, {"text": "C... | |
hard | 14 | The figure illustrates a dual-classifier framework for learning visually grounded predictions from noisy human-centric annotations. The global layout is left-to-right, beginning with an input image of bananas labeled with concepts 'Banana' and 'Yellow'. This image feeds into a ConvNet, whose fc7 layer outputs feature r... | 2,048 | 512 | 1512.06974 | 2015 | 1512.06974/1512.06974_5_img_SINGLE.png | [{"name": "Input Image", "description": "Example photograph plus word set w ∈ {Banana, Yellow}", "importance": "primary"}, {"name": "ConvNet", "description": "Convolutional neural network; fc7 feature vector is forwarded to later modules", "importance": "primary"}, {"name": "Visual Presence Classifier", "description": ... | [{"source": "Input Image", "target": "ConvNet", "label": null, "importance": "primary"}, {"source": "ConvNet", "target": "Visual Presence Classifier", "label": "fc7", "importance": "primary"}, {"source": "ConvNet", "target": "Relevance Classifier", "label": "fc7", "importance": "primary"}, {"source": "Visual Presence C... | [{"text": "w ∈ {Banana, Yellow}", "location": "beneath the input image", "importance": "primary"}, {"text": "fc7", "location": "on both arrows from ConvNet to the two classifiers", "importance": "secondary"}, {"text": "Linear", "location": "first small box inside both classifiers", "importance": "secondary"}, {"text": ... | |
hard | 15 | The figure illustrates a multi-stream deep learning framework for video classification, structured into three parallel processing streams: spatial, motion, and audio. The global layout is vertically stacked, with each stream occupying a distinct horizontal section, all converging toward a final prediction module on the... | 896 | 1,088 | 1509.06086 | 2015 | 1509.06086/1509.06086_0_img_SINGLE.png | [{"name": "Input Video Clip", "description": "Stack of consecutive video frames that serves as the common raw input to all three processing streams.", "importance": "primary"}, {"name": "Individual frame", "description": "A single RGB frame extracted from the video clip, used by the spatial stream.", "importance": "pri... | [{"source": "Input Video Clip", "target": "Individual frame", "label": null, "importance": "primary"}, {"source": "Input Video Clip", "target": "Stacked optical flow", "label": null, "importance": "primary"}, {"source": "Input Video Clip", "target": "Audio spectrogram", "label": null, "importance": "primary"}, {"source... | [{"text": "Individual frame", "location": "Top-left of spatial stream box", "importance": "secondary"}, {"text": "Spatial ConvNet", "location": "Inside top box above ConvNet layers", "importance": "secondary"}, {"text": "Spatial LSTM", "location": "Below top LSTM sequence", "importance": "secondary"}, {"text": "Stacked... | |
hard | 16 | The figure illustrates a deep learning framework integrating the radius-margin bound as a regularizer to enhance classification generalization. The global layout is horizontal and sequential, depicting two stages of feature space transformation separated by a central dashed box representing the deep neural network arch... | 1,728 | 576 | 1512.01642 | 2015 | 1512.01642/1512.01642_3_img_SINGLE.png | [{"name": "Left Feature Sphere (iteration k-1)", "description": "Circular region depicting the data distribution before transformation; contains red triangles, green circles, a dashed decision boundary and its margin, inward radial arrows, and the annotation R_{k-1}.", "importance": "primary"}, {"name": "Transformation... | [{"source": "Left Feature Sphere (iteration k-1)", "target": "Transformation Block", "label": null, "importance": "primary"}, {"source": "Transformation Block", "target": "Right Feature Sphere (iteration k)", "label": null, "importance": "primary"}] | [{"text": "R_{k-1}", "location": "Inside left sphere, below the decision boundary", "importance": "primary"}, {"text": "R_k", "location": "Inside right sphere, below the decision boundary", "importance": "primary"}, {"text": "\\frac{2}{\\|\\mathbf{w}_{k-1}\\|}", "location": "Above left sphere, near the top margin indic... | |
hard | 17 | The figure illustrates the architecture of a Scale-Aware Feature R-CNN (SAF R-CNN), designed for object detection with improved performance across different object scales. The global layout is left-to-right, starting with an input image on the far left, progressing through a shared backbone network, and then branching ... | 1,344 | 768 | 1510.08160 | 2015 | 1510.08160/1510.08160_2_img_SINGLE.png | [{"name": "Input Image", "description": "Original scene picture with marked regions of interest.", "importance": "primary"}, {"name": "Feature Extraction Network", "description": "Stacked convolutional layers that turn the input image into multi-scale feature maps (channels 64→128→256).", "importance": "primary"}, {"na... | [{"source": "Input Image", "target": "Feature Extraction Network", "label": null, "importance": "primary"}, {"source": "Feature Extraction Network", "target": "Large-size Sub-network", "label": null, "importance": "primary"}, {"source": "Feature Extraction Network", "target": "Small-size Sub-network", "label": null, "i... | [{"text": "input", "location": "top-left near the image", "importance": "secondary"}, {"text": "64", "location": "under first feature cube", "importance": "secondary"}, {"text": "128", "location": "under second feature cube", "importance": "secondary"}, {"text": "256", "location": "under third feature cube", "importanc... | |
hard | 18 | The figure illustrates the unrolled architecture of a CST-VAE (Conditional Style Transfer Variational Autoencoder) network for processing multiple image layers, specifically shown for two layers (L₁ and L₂). The global layout is horizontal and sequential, with each layer processed independently but sharing a common str... | 1,280 | 768 | 1511.06362 | 2015 | 1511.06362/1511.06362_6_img_SINGLE.png | [{"name": "Input image / initial residual", "description": "Top-left gray node labeled Δ1 = x, serving as the initial input to the first layer branch."}, {"name": "Residual propagation nodes", "description": "Top-row gray/white nodes for Δ1 and Δ2, carrying the residual or remaining image information across layers."}, ... | [{"source": "Input image / initial residual", "target": "Residual propagation nodes", "label": "residual flow across layers"}, {"source": "Residual propagation nodes", "target": "Layer 1 pose encoder", "label": "input to pose encoder"}, {"source": "Residual propagation nodes", "target": "Layer 1 content encoder", "labe... | [{"text": "Δ1 = x", "location": "top left above first residual node"}, {"text": "Δ2", "location": "top center above second residual node"}, {"text": "STN(Δ1, T1^-1)", "location": "diagonal label between first residual/input and layer 1 content encoder"}, {"text": "STN(C1, T1)", "location": "below layer 1 decoders above... | |
hard | 19 | The figure illustrates the concept of generating multi-scale part proposals from an object proposal using a Convolutional Neural Network (CNN), specifically focusing on the CONV5 layer. The global layout is divided into two main sections: on the left, a 3D representation of an object proposal (labeled 'Object proposal ... | 1,408 | 704 | 1504.04943 | 2015 | 1504.04943/1504.04943_3_img_SINGLE.png | [{"name": "Object proposal (CNN input)", "description": "The 224×224 image patch fed into the CNN; shown in perspective on the left.", "importance": "primary"}, {"name": "CONV5 layer (N×N grid)", "description": "Feature-map grid produced by the CNN; illustrated as an N-by-N square on the right.", "importance": "primary... | [{"source": "Receptive field of P1 (139×139)", "target": "P1", "label": null, "importance": "primary"}, {"source": "Receptive field of P2 (155×155)", "target": "P2", "label": null, "importance": "primary"}] | [{"text": "224", "location": "Left edge of input image", "importance": "secondary"}, {"text": "224", "location": "Bottom edge of input image", "importance": "secondary"}, {"text": "Receptive field of P1", "location": "Inside blue rectangle", "importance": "secondary"}, {"text": "139", "location": "Two instances inside ... | |
hard | 20 | The figure illustrates an end-to-end neural network architecture for a sequence prediction task, likely involving visual input and sequential output generation. The global layout is hierarchical and sequential, structured from top to bottom and left to right. At the top, there is a photographic image of a dining area w... | 1,152 | 832 | 1505.01121 | 2015 | 1505.01121/1505.01121_1_img_SINGLE.png | [{"name": "Input Image", "description": "Photograph that serves as visual input to the network.", "importance": "primary"}, {"name": "CNN Encoder", "description": "Trapezoid-shaped block labelled “CNN” that extracts visual features from the image.", "importance": "primary"}, {"name": "LSTM Cell 1", "description": "Firs... | [{"source": "Input Image", "target": "CNN Encoder", "label": null, "importance": "primary"}, {"source": "CNN Encoder", "target": "LSTM Cell 1", "label": "x", "importance": "primary"}, {"source": "CNN Encoder", "target": "LSTM Cell 2", "label": "x", "importance": "primary"}, {"source": "CNN Encoder", "target": "LSTM Cel... | [{"text": "x", "location": "on the arrows broadcasting CNN features to LSTMs", "importance": "primary"}, {"text": "q_{n-1}", "location": "inside the left green query box", "importance": "primary"}, {"text": "q_n", "location": "inside the second green query box", "importance": "primary"}, {"text": "LSTM", "location": "i... | |
hard | 21 | The figure illustrates the architecture of an HT (Hierarchical Transformation) model, which is a convolutional arithmetic circuit designed to implement hierarchical decomposition. The global layout is a left-to-right sequential pipeline, starting from an input X and progressing through multiple hidden layers to a final... | 1,920 | 512 | 1509.05009 | 2015 | 1509.05009/1509.05009_1_img_SINGLE.png | [{"name": "Input X", "description": "Input cube containing the sample x_i", "importance": "primary"}, {"name": "Representation", "description": "Feature-extraction block computing rep(i,d)=f_{θ_d}(x_i)", "importance": "primary"}, {"name": "Conv 0 (1x1)", "description": "First hidden-layer 1×1 convolution computing conv... | [{"source": "Input X", "target": "Representation", "label": null, "importance": "primary"}, {"source": "Representation", "target": "Conv 0 (1x1)", "label": null, "importance": "primary"}, {"source": "Conv 0 (1x1)", "target": "Pooling 0", "label": null, "importance": "primary"}, {"source": "Pooling 0", "target": "Interm... | [{"text": "input X", "location": "above the leftmost cube", "importance": "primary"}, {"text": "x_i", "location": "inside the green cube", "importance": "secondary"}, {"text": "representation", "location": "above the second cube", "importance": "primary"}, {"text": "rep(i,d)=f_{θ_d}(x_i)", "location": "below the repres... | |
hard | 22 | The figure illustrates a hybrid neural network architecture combining a scene feature extraction module with a modified Long Short-Term Memory (LSTM) unit, referred to as SF-LSTM (Scene-Factorized LSTM). The global layout is divided into two main parts: on the left, a scene vector extraction pipeline; on the right, the... | 1,536 | 640 | 1506.06272 | 2015 | 1506.06272/1506.06272_9_img_SINGLE.png | [{"name": "Image Input", "description": "A baseball scene image serving as raw visual input.", "importance": "primary"}, {"name": "Places-205 CNN", "description": "Convolutional neural network that extracts high-level scene features.", "importance": "primary"}, {"name": "2-layer MLP", "description": "Two-layer multilay... | [{"source": "Image Input", "target": "Places-205 CNN", "label": null, "importance": "primary"}, {"source": "Places-205 CNN", "target": "2-layer MLP", "label": null, "importance": "primary"}, {"source": "2-layer MLP", "target": "Scene Vector s", "label": "scene vector", "importance": "primary"}, {"source": "Scene Vector... | [{"text": "Places-205 CNN", "location": "inside green rounded rectangle", "importance": "secondary"}, {"text": "2-layer MLP", "location": "inside green rounded rectangle", "importance": "secondary"}, {"text": "s", "location": "inside small circle after MLP", "importance": "secondary"}, {"text": "SF-LSTM", "location": "... | |
hard | 23 | The figure illustrates the overall procedure of a discriminative and generative tracking algorithm that leverages a pre-trained CNN for feature extraction and target localization. The global layout is structured as a two-row workflow: the top row represents the discriminative component involving sample classification a... | 1,728 | 576 | 1502.06796 | 2015 | 1502.06796/1502.06796_1_img_SINGLE.png | [{"name": "Input frame", "description": "Current video frame containing the target object, displayed with several candidate bounding boxes.", "importance": "primary"}, {"name": "Samples", "description": "Set of cropped image regions (candidate patches) extracted from the input frame.", "importance": "primary"}, {"name"... | [{"source": "Input frame", "target": "Samples", "label": null, "importance": "primary"}, {"source": "Samples", "target": "Pre-trained CNN", "label": "forward", "importance": "primary"}, {"source": "Pre-trained CNN", "target": "Sample features", "label": null, "importance": "primary"}, {"source": "Sample features", "tar... | [{"text": "forward", "location": "on arrow between Samples and Pre-trained CNN", "importance": "secondary"}, {"text": "backward", "location": "on arrow between Target-specific features and Pre-trained CNN; also on arrow between Pre-trained CNN and Target-specific saliency map", "importance": "secondary"}, {"text": "×",... | |
hard | 24 | The figure illustrates a deep learning framework for image segmentation, specifically designed for semantic labeling tasks such as identifying regions like 'sea', 'sky', and 'sand' in an input image. The global layout is structured as a two-tiered pipeline: a top-level flow showing the high-level processing stages, and... | 1,408 | 704 | 1512.03131 | 2015 | 1512.03131/1512.03131_15_img_SINGLE.png | [{"name": "Input image with RGB-location channels", "description": "Stacked multi-channel input patch representing the image and location information."}, {"name": "Sampled local patches/regions", "description": "Highlighted square regions on the input image indicating local areas used for feature extraction."}, {"name"... | [{"source": "Input image with RGB-location channels", "target": "Truncated Convolutional Neural Network", "label": "input to feature extractor"}, {"source": "Truncated Convolutional Neural Network", "target": "Intermediate feature map stack", "label": "produces feature maps"}, {"source": "Intermediate feature map stack... | [{"text": "Metrics", "location": "top center"}, {"text": "Truncated Convolutional Neural Network", "location": "center left"}, {"text": "Global Belief", "location": "upper right branch"}, {"text": "Local Belief", "location": "middle right branch"}, {"text": "Softmax", "location": "inside local belief branch"}, {"text":... | |
hard | 25 | The figure illustrates the architecture of a multimodal caption generation model, composed of three main components: a Lexical Classifier, a Caption Model (which includes a Multimodal Unit), and a Language Model. These components are arranged horizontally from left to right, each enclosed in a rounded rectangular box w... | 1,152 | 832 | 1511.05284 | 2015 | 1511.05284/1511.05284_12_img_SINGLE.png | [{"name": "Input Image", "description": "Photographic image that serves as the visual input to both the Lexical Classifier and (through an adapter) the Caption Model.", "importance": "primary"}, {"name": "Lexical Classifier", "description": "Left-hand module that extracts a generic visual feature vector from the input ... | [{"source": "CNN", "target": "Classification Layer", "label": null, "importance": "primary"}, {"source": "Classification Layer", "target": "f_I", "label": null, "importance": "primary"}, {"source": "Input Image", "target": "Image Adapter Funnel", "label": null, "importance": "primary"}, {"source": "Image Adapter Funnel... | [{"text": "Weights pre-trained with lexical classifier", "location": "Legend – yellow square item", "importance": "secondary"}, {"text": "Weights pre-trained with language model", "location": "Legend – blue square item", "importance": "secondary"}, {"text": "Weights learned only when training caption model", "location"... | |
hard | 26 | The figure illustrates a deep unstructured convolutional activation framework for scene classification. The global layout is a horizontal workflow divided into two main rows: the top row represents the feature extraction and representation pipeline, while the bottom row details the classification process. A vertical da... | 1,792 | 576 | 1506.05196 | 2015 | 1506.05196/1506.05196_1_img_SINGLE.png | [{"name": "Office Scene Input", "description": "Initial set of input images representing an office-like scene, shown with a question mark indicating unknown class.", "importance": "primary"}, {"name": "Mid-level Patches", "description": "Cloud-like group of small image patches extracted from the input scene (Sec. 3.1).... | [{"source": "Office Scene Input", "target": "Mid-level Patches", "label": null, "importance": "primary"}, {"source": "Mid-level Patches", "target": "Convolutional Neural Network Activations", "label": null, "importance": "primary"}, {"source": "Convolutional Neural Network Activations", "target": "Scene Representative ... | [{"text": "Office Scene ?", "location": "above the input images", "importance": "primary"}, {"text": "Mid-level Patches (Sec. 3.1)", "location": "above the cloud of patches", "importance": "secondary"}, {"text": "Convolutional Neural Network Activations (Sec. 3.2)", "location": "above the CNN block diagram", "importanc... |
End of preview. Expand in Data Studio
SciFormaBench-2K: Full-Reference Benchmark Backup
This repository is the complete-reference backup of SciFormaBench-2K, the official evaluation benchmark for SciForma. It contains all 2,000 reference images, prompts, target resolutions, provenance fields, and Component/Arrow/Text structural rubrics.
Features
- 🖼️ Complete reference set. All 2,000 ground-truth scientific methodology diagrams are embedded in the Parquet files.
- 🧩 Official benchmark splits. Simple/Medium/Hard contain 500/900/600 samples with the original benchmark indices and ordering.
- 🎯 Axis-specific evaluation. Each sample includes independently verified Component, Arrow, and Text inventories.
- 🔗 Source-linked records. Each row retains its arXiv ID, year, source image path, prompt, and target dimensions.
- ✅ Complete text rubrics.
rubric_textis populated from the officialstage1_abstraction.text_labels; 1,873 rows contain non-empty text-label inventories.
Usage
import json
from datasets import load_dataset
dataset = load_dataset(
"LoYuXrqw/SciFormaBench-2K",
"hard", # simple / medium / hard
split="test",
token=True,
)
sample = dataset[0]
sample["gt_image"].show()
components = json.loads(sample["rubric_components"])
arrows = json.loads(sample["rubric_arrows"])
text_labels = json.loads(sample["rubric_text"])
print(sample["index"], len(components), len(arrows), len(text_labels))
Set HF_TOKEN or log in with hf auth login before loading this repository.
Evaluation
Use the official SciForma evaluation pipeline:
python eval/eval_benchmark.py \
--gen_dir ./results/my_model \
--output_dir ./eval_results/my_model \
--hf_benchmark LoYuXrqw/SciFormaBench-2K \
--levels hard \
--num_retest 2 \
--deployment_name gpt-5.4
Use --levels simple medium hard for the complete benchmark. The paper uses
GPT-5.4 as the judge; GPT-4o is the public alternative. The official protocol
uses axis-specific calls with split_dims=True and averages two evaluation
rounds with num_retest=2.
License
This full-reference repository is maintained for internal preservation and evaluation. It does not provide a new redistribution license for the embedded figures. Each image remains governed by its original source-paper license.
- Downloads last month
- 34