Datasets:
StepCount-Extreme
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:1471
- 数量范围:51–500
- 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=132,fsc147_test=1,fscd_147=300,fscd_lvis=279,sku110k=300,visdrone_det=159 - 领域:
aerial=159,cluttered_multiclass=279,crowd=300,everyday=301,generic_dense=132,retail=300 - 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
export HF_TOKEN=<your-approved-token>
hf download SI-Lab/StepCount-Extreme --type dataset --local-dir ./StepCount-Extreme
python ./StepCount-Extreme/scripts/materialize_eval_json.py
物化后使用 ./StepCount-Extreme/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
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-Extreme/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 501 \
--base_dot_radius 8 \
--base_font_size 16 \
--pass_k 1
max_new_tokens=8192,greedy decoding(do_sample=false)。
Qwen2.5-VL-7B direct-count 冻结配置
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-Extreme/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-3.5-Flash | 1471 | 1470 | 5.23% | 26.542 | 11.35% | 45.07% | — | — |
| Gemini-3.1-Pro | 1471 | 1470 | 4.35% | 29.017 | 10.2% | 36.3% | — | — |
| LocateAnything-3B | 1471 | 1471 | 2.31% | 109.26 | 4.62% | 12.24% | 21.46% | 24.1% |
| Gemini-2.5-Pro | 1471 | 1401 | 2.24% | 75.751 | 4.62% | 20.6% | — | — |
| GPT-5.4 | 1471 | 1471 | 1.77% | 41.158 | 3.67% | 25.9% | — | — |
| Claude-4.6-Sonnet | 1471 | 1467 | 1.5% | 43.108 | 4.28% | 19.85% | — | — |
| Claude-Opus-4-7 | 1471 | 1449 | 1.16% | 58.034 | 3.81% | 15.84% | — | — |
| Qwen3.6-27B | 1471 | 1337 | 0.95% | 60.755 | 2.18% | 11.62% | — | — |
| Qwen3-VL-8B-Instruct | 1471 | 1430 | 0.61% | 48.276 | 1.63% | 12.71% | — | — |
| Qwen2.5-VL-7B-bf16-0714 | 1471 | 1463 | 0.54% | 51.362 | 1.56% | 10.67% | — | — |
| StepCount-SFT-1M-bf16-0714 | 1471 | 1471 | 0.48% | 229.859 | 1.36% | 5.1% | 30.89% | 47.04% |
| Qwen3.5-9B | 1471 | 1430 | 0.41% | 67.289 | 1.5% | 8.36% | — | — |
| Qwen2.5-VL-7B | 1471 | 1471 | 0.27% | 56.4 | 1.22% | 8.02% | — | — |
| Qwen3-VL-32B-Instruct | 1471 | 1471 | 0.14% | 55.657 | 0.41% | 3.94% | — | — |
| Molmo-7B-D | 1471 | 1471 | 0.0% | 108.179 | 0.0% | 0.0% | 1.48% | 1.79% |
| Molmo2-8B | 1471 | 1471 | 0.0% | 109.095 | 0.0% | 0.0% | — | — |
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。
- Downloads last month
- 26