| --- |
| license: other |
| license_name: mixed-source-research-non-commercial |
| pretty_name: StepCount-Dense |
| task_categories: |
| - visual-question-answering |
| tags: |
| - counting |
| - object-counting |
| - visual-grounding |
| - benchmark |
| - gated |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: eval/test.json |
| --- |
| |
| # StepCount-Dense |
|
|
| StepCount-Bench v3 的 eval-ready 子集。该仓库用于研究与非商业评测，下载采用 **manual gated approval**。 |
|
|
| ## 数据概览 |
|
|
| - 样本数：**1229** |
| - 数量范围：**11–50** |
| - 每个样本包含：final count、归一化 `gt_points`、可选 `gt_bbox`、domain、scene level、ID/OOD category split、source provenance。 |
| - 数据源：`crowdhuman`=300, `dense200`=67, `fsc147_test`=1, `fscd_147`=300, `fscd_lvis`=300, `sku110k`=9, `visdrone_det`=252 |
| - 领域：`aerial`=252, `cluttered_multiclass`=300, `crowd`=300, `everyday`=301, `generic_dense`=67, `retail`=9 |
| - 坐标契约：`gt_points=[x,y]`、`gt_bbox=[x1,y1,x2,y2]`，均归一化到 `[0,1]`。 |
| - 完整机器可读榜单：`metadata/leaderboard.json` |
| - 冻结评测协议：`metadata/eval_config.json` |
|
|
| ## 下载与直接评测 |
|
|
| ```bash |
| export HF_TOKEN=<your-approved-token> |
| hf download SI-Lab/StepCount-Dense --type dataset --local-dir ./StepCount-Dense |
| python ./StepCount-Dense/scripts/materialize_eval_json.py |
| ``` |
|
|
| 物化后使用 `./StepCount-Dense/eval/test.absolute.json`，其 `image_path` 为当前机器绝对路径，可直接传给现有 `eval.py` 与 `eval_oneshot_pointing.py`。 |
|
|
| ### StepCount 冻结配置 |
|
|
| ```bash |
| export STEPCOUNT_DTYPE=bfloat16 |
| export STEPCOUNT_ATTN_IMPL=sdpa |
| python3 eval/scripts/eval.py \ |
| --model_path /path/to/StepCount-model \ |
| --eval_dataset_path /absolute/path/StepCount-Dense/eval/test.absolute.json \ |
| --output_dir /path/to/eval-output \ |
| --output_images_dir /path/to/eval-output/images \ |
| --output_json_path /path/to/results.json \ |
| --max_rounds 51 \ |
| --base_dot_radius 10 \ |
| --base_font_size 20 \ |
| --pass_k 1 |
| ``` |
|
|
| `max_new_tokens=8192`，greedy decoding（`do_sample=false`）。 |
|
|
| ### Qwen2.5-VL-7B direct-count 冻结配置 |
|
|
| ```bash |
| python3 eval/scripts/eval_oneshot_pointing.py \ |
| --model_type qwen_auto \ |
| --model_path /path/to/Qwen2.5-VL-7B-Instruct \ |
| --bench_json /absolute/path/StepCount-Dense/eval/test.absolute.json \ |
| --output_json /path/to/results.json |
| ``` |
|
|
| Prompt：原始问题后追加 `Carefully count and answer with a single integer only.`；bf16、SDPA 优先、greedy、`max_new_tokens=128`，仅 standalone integer 视为有效预测。 |
|
|
| ## 指标定义 |
|
|
| - **Exact Match (EM)**：`pred == gt` 的样本比例；无效预测按错误计入完整分母。 |
| - **MAE**：`mean(abs(pred-gt))`；仅在 valid prediction 上计算，同时报告 `valid_n`。 |
| - **Within-1**：`abs(pred-gt) <= 1`；使用完整分母。 |
| - **Within-10%**：`abs(pred-gt)/gt <= 0.1`；使用完整分母。 |
| - **Point Precision / Recall / F1**：在归一化坐标中以阈值优先 Hungarian matching 计算；没有输出点的样本以 0 计入。 |
| - **Coverage@N**：前 N 个预测点对 GT 的覆盖率。 |
| - **Double-Count**：超过可匹配 GT 数量的重复/额外点比例。 |
|
|
| ## Canonical leaderboard |
|
|
| | Model | n | valid | EM ↑ | MAE ↓ | Within-1 ↑ | Within-10% ↑ | Point F1 ↑ | Coverage@N ↑ | |
| |---|---:|---:|---:|---:|---:|---:|---:|---:| |
| | Gemini-3.1-Pro | 1229 | 1229 | 23.19% | 7.304 | 36.45% | 46.38% | — | — | |
| | Gemini-3.5-Flash | 1229 | 1228 | 22.54% | 6.945 | 37.67% | 48.09% | — | — | |
| | LocateAnything-3B | 1229 | 1229 | 16.52% | 25.731 | 25.39% | 30.76% | 44.41% | 45.86% | |
| | GPT-5.5 | 1229 | 1227 | 14.65% | 12.488 | 28.07% | 36.13% | — | — | |
| | Molmo-7B-D | 1229 | 1229 | 14.08% | 17.541 | 23.35% | 26.12% | 33.88% | 34.77% | |
| | Gemini-2.5-Pro | 1229 | 1227 | 12.53% | 21.581 | 27.58% | 36.53% | — | — | |
| | StepCount-SFT-1M-bf16-0714 | 1229 | 1229 | 11.15% | 12.772 | 20.67% | 27.58% | 49.85% | 58.39% | |
| | Claude-4.6-Sonnet | 1229 | 1229 | 9.28% | 9.573 | 21.16% | 28.4% | — | — | |
| | Claude-Opus-4-7 | 1229 | 1221 | 8.79% | 10.617 | 19.45% | 24.74% | — | — | |
| | Qwen3.5-9B | 1229 | 1197 | 8.79% | 9.431 | 20.26% | 25.39% | — | — | |
| | Qwen3-VL-32B-Instruct | 1229 | 1229 | 8.62% | 10.25 | 19.2% | 23.19% | — | — | |
| | GPT-5.4 | 1229 | 1229 | 8.38% | 9.037 | 22.78% | 30.35% | — | — | |
| | Qwen3.6-27B | 1229 | 1137 | 7.24% | 10.843 | 18.14% | 23.11% | — | — | |
| | StepCount-SFT-1M | 1229 | 1229 | 7.16% | 13.254 | 14.08% | 18.06% | 33.1% | 43.11% | |
| | Qwen2.5-VL-7B-bf16-0714 | 1229 | 1225 | 6.83% | 10.311 | 19.69% | 24.74% | — | — | |
| | StepCount-7b-v12-h20 | 1229 | 1229 | 6.43% | 11.886 | 14.08% | 19.2% | 38.86% | 48.37% | |
| | Qwen2.5-VL-7B | 1229 | 1229 | 6.35% | 10.072 | 19.85% | 24.25% | — | — | |
| | Qwen2.5-VL-7B-point | 1229 | 1229 | 6.18% | 18.015 | 15.13% | 19.12% | 44.89% | 52.97% | |
| | Qwen3-VL-8B-Instruct | 1229 | 1201 | 5.94% | 11.188 | 16.03% | 20.18% | — | — | |
| | StepCount-7B_v33_step40 | 1229 | 1229 | 5.53% | 13.621 | 13.26% | 16.6% | 35.27% | 45.46% | |
| | StepCount-v34A-step101 | 1229 | 1229 | 5.37% | 14.018 | 13.59% | 16.6% | 31.87% | 41.61% | |
| | StepCount-v35-step101 | 1229 | 1229 | 5.21% | 13.002 | 12.37% | 15.54% | 34.04% | 44.3% | |
| | StepCount-v35-step80 | 1229 | 1229 | 5.21% | 13.002 | 12.37% | 15.54% | 34.04% | 44.3% | |
| | StepCount-SFT | 1229 | 1229 | 4.88% | 13.78 | 12.86% | 16.92% | 33.8% | 44.04% | |
| | StepCount-v34A-step80 | 1229 | 1229 | 4.72% | 14.041 | 12.86% | 15.87% | 31.71% | 41.52% | |
|
|
| ## 许可与访问限制 |
|
|
| 这是 mixed-source benchmark，不提供统一商业许可。访问者必须同时遵守各原始数据集条款： |
|
|
| - **CrowdHuman**：仅限学术/研究用途，原始条款可能限制再分发。 |
| - **VisDrone**：CC BY-NC-SA 3.0，要求署名、非商业、相同方式共享。 |
| - **FSCD-147 / FSCD-LVIS / FSC-147 / SKU110K / Dense200**：遵循各自原始项目许可与引用要求。 |
|
|
| Manual gate 不授予任何超出源数据许可的权利。若源数据所有者要求移除内容，请联系 SI-Lab。 |
|
|
| ## 数据完整性 |
|
|
| 上传前验证：ID 唯一、图片存在、`answer == len(gt_points)`、point/bbox 坐标合法。`checksums.sha256` 可用于下载后逐文件校验。 |
|
|
