Clarify benchmark scope and add context-scaling documentation

README.md

@@ -11,18 +11,39 @@ tags:
 
 # MindView-Bench
 
-MindView benchmarks language-guided target-view generation from multiple observations of the same indoor scene. **MindView-Bench** is the evaluation-data repository for *MindView: Benchmarking Generative Multi-View Spatial Intelligence*.
+MindView evaluates **language-guided target-view generation**: a model uses observations of an indoor scene and a language query to infer the requested viewpoint and generate a view consistent with that scene. **MindView-Bench** is the evaluation-data repository for *MindView: Benchmarking Generative Multi-View Spatial Intelligence*.
 
 **Release status:** This repository currently contains documentation only. Dataset files and image assets have not been uploaded. Release packaging and access instructions are being prepared.
 
-## Benchmark specification
+## Official benchmark
 
-- **1,500 aligned scene–query–target samples** from **513 indoor scenes**.
-- **Six context images** per sample in the standard setting.
-- Nested **C3, C6, and C10** context settings, yielding **4,500 experiment records** across the same 1,500 samples.
-- Seven tasks: Translation, Rotation, Composition, Relational Object Grounding, Target-Facing Reorientation, Object-Centric Repositioning, and Multi-Cue Viewpoint Inference.
+The official benchmark comprises **1,500 distinct scene–query–target samples from 513 indoor scenes**, drawn from ScanNet, ScanNet++, and ARKitScenes. Each sample provides ordered RGB observations and a language query, with a separate held-out real target for evaluation.
 
-The benchmark metadata identifies the scene, task, language instruction, target frame, and ordered context frames. RGB images and sensor-depth resources are separate assets; the geometric evaluation workflow requires metric depth. Images and scene resources originate from ScanNet, ScanNet++, and ARKitScenes.
+Its seven tasks form two complementary families:
+
+| Family | Tasks |
+|---|---|
+| Geometry-dominant | Translation; Rotation; Composition |
+| Semantics-dominant | Relational Object Grounding; Target-Facing Reorientation; Object-Centric Repositioning; Multi-Cue Viewpoint Inference |
+
+Geometry-dominant tasks specify camera motions from an anchor. Semantics-dominant tasks require viewpoint inference from referred objects, spatial relations, and reference-frame cues.
+
+## Evaluation
+
+Evaluation measures **DINO-I, Semantic, Spatial, Quality, VGGT-RScore, and VGGT-TScore**. Each component is averaged within tasks and then equally across the seven tasks. **Overall** is the unweighted mean of these six task-macro scores under the official benchmark protocol.
+
+The release metadata identifies the scene, task, instruction, target frame, and ordered context frames. RGB images are the visual generation inputs; sensor-depth resources support geometric scoring. Asset manifests, checksums, and loading instructions will be documented with the data release.
+
+## Repository layout
+
+```text
+MindView-Bench/
+├── README.md
+└── context_scaling/
+    └── README.md
+```
+
+See [Context-scaling experiments](context_scaling/README.md) for the supplementary paired analysis of input-context budgets. Its configurations and reporting protocol are described separately from the official benchmark.
 
 ## Related resources
 
@@ -30,12 +51,12 @@ Collection: [MindView](https://huggingface.co/collections/wangyanbo/mindview-6ac
 
 | Resource | Purpose |
 |---|---|
-| [MindView-Bench](https://huggingface.co/datasets/wangyanbo/MindView-Bench) | Frozen evaluation data and release metadata |
-| [MindView-Train](https://huggingface.co/datasets/wangyanbo/MindView-Train) | Six-view supervised training data |
+| [MindView-Bench](https://huggingface.co/datasets/wangyanbo/MindView-Bench) | Official evaluation benchmark and supplementary experiment documentation |
+| [MindView-Train](https://huggingface.co/datasets/wangyanbo/MindView-Train) | Supervised training data |
 | [MindView-BAGEL-SFT](https://huggingface.co/wangyanbo/MindView-BAGEL-SFT) | Fine-tuned BAGEL EMA checkpoint |
 | [GitHub project repository](https://github.com/Yanbo-23/MindView) | Project code and usage documentation |
 
-Code and usage documentation are being prepared for publication at the project repository. Loading instructions, asset manifests, and checksums will be documented with the data release.
+Code and usage documentation are being prepared for publication at the project repository.
 
 ## Access and terms
 

context_scaling/README.md

@@ -0,0 +1,39 @@
+# Context-scaling experiments
+
+This directory documents supplementary experiments that vary the number of input observations on the **same 1,500 samples** used by [MindView-Bench](../README.md). These experiments study evidence utilization under reduced and expanded context.
+
+**Release status:** This directory currently contains documentation only. Experiment metadata and image assets have not been uploaded. Their file layout and loading instructions will be documented with the release.
+
+## Paired configurations
+
+| Configuration | Input context images | Role in this analysis |
+|---|---:|---|
+| **C3** | 3 | Reduced context |
+| **C6** | 6 | Reference configuration |
+| **C10** | 10 | Expanded context |
+
+Each sample is evaluated once in each configuration. The three configurations therefore give **4,500 configuration records across the same 1,500 distinct samples**. The sample count remains 1,500; the configuration records do not add new independent benchmark samples.
+
+The pairing preserves the sample identity, scene, task, language query, and held-out target. Contexts are strictly nested ordered subsequences: **C3 ⊂ C6 ⊂ C10**. Shared views retain their order, the target is excluded from every input context, and any required anchor remains first.
+
+## Generation inputs and scoring references
+
+A generator receives the context images for its selected configuration and the language query.
+
+For **VGGT-based geometric scoring**, the evaluator uses the same ten ordered reference views from the sample's ten-context set across configurations and generators. It jointly processes these references, the real target, and the generated image. Sensor depths from the real reference frames and target calibrate translation error to meters.
+
+These fixed references and sensor-depth resources belong to evaluation. They do not increase the context supplied to a generator in the reduced or reference configuration, and the held-out target is used only for scoring.
+
+## Reporting
+
+Report scores separately for C3, C6, and C10, with changes computed on matched samples. Within each configuration, the six components use the same seven-task macro-averaging protocol as the benchmark.
+
+**Official benchmark Overall follows the official benchmark protocol. It is not an average across the three context configurations.** Cross-configuration summaries, when reported, should be labeled as supplementary context-scaling results and retain the 1,500-sample denominator.
+
+## Project resources
+
+See the [benchmark card](../README.md) for the official benchmark definition, related training data and model, release status, access terms, and project contacts.
+
+Collection: [MindView](https://huggingface.co/collections/wangyanbo/mindview-6acf6b7a4a2e25efd51989f5). Project repository: [GitHub](https://github.com/Yanbo-23/MindView).
+
+Code and usage documentation are being prepared for publication at the project repository.