Spaces:
Sleeping
Sleeping
File size: 3,305 Bytes
b9784cd 68fa834 b9784cd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | # 📘 MMScan Hierarchical VIsual Grounding Challenge
## 🔍 Challenge Introduction
**Hierarchical Visual Grounding (HVG) Task in the MMScan Benchmark**:
This task evaluates a model’s ability to perform visual grounding at multiple levels of granularity — from region to object-level, and from single-target localization to inter-targets localization. Given a natural language description, models are expected to accurately locate the corresponding object(s) within the 3D scenes, reflecting comprehensive spatial and attribute-level understanding.
- **Overview**: You can refer to this [website](https://neurips.cc/virtual/2024/poster/97429) for an overview and our [paper](https://arxiv.org/abs/2406.09401) for more details.
- **Challenge Data and Codebase**: The challenge dataset includes:
- **Training set**: Language prompts + ground-truth bounding boxes
- **Validation set**: Language prompts + ground-truth bounding boxes
- **Test set**: Language prompts only (no ground truth provided)
Follow the [instructions](https://github.com/OpenRobotLab/EmbodiedScan/tree/mmscan) to get familiar with data organization and MMScan APIs. All the code for MMScan is available [here](https://github.com/OpenRobotLab/EmbodiedScan/tree/mmscan).
- **Evaluation Metrics**: For the visual grounding task, our evaluator computes multiple metrics including AP@0.25 (Average Precision), gTop-1@0.25, and gTop-3@0.25 where the gTop-k metric is an expanded metric that generalizes the traditional Top-k metric, offering superior flexibility and interpretability compared to traditional ones when oriented towards multi-target grounding.
- **Contact**: For any questions related to the HVG challenge, feel free to reach out to [**Jingli Lin**](linjingli166@gmail).
---
## 📝 How to Participate
To register for the challenge, please contact us via [**Google Mail**](linjingli166@gmail) and include the following information:
- A **self-chosen username** (this will be shown on the leaderboard)
- A **login password**
- Your **team or institution name**
- A brief statement on your **motivation for participating**
> 📌 **Submission limit**: Each user is allowed a **maximum of 5 submissions per day**.
---
## 🚀 Submission Guidelines
- Your submission should be a **dictionary**, where each key is a **sample ID** from the test split.
- For each sample, provide:
- `pred_bboxes`: a list of predicted bounding boxes
- `scores`: the corresponding confidence scores
- An expected result is:
```python
{
'VG_Inter_Space_OO__1mp3d_0009_region0__55'(sample ID):
{
'pred_bboxes'(list, 100*9): [[...],...]
'scores'(list, 100): [...]
}
...
}
```
> 💡 **Note**: The bounding boxes do **not** need to be sorted by confidence.
- ⛔ **Limit the number of predicted boxes to 100 per sample**. If your submission contains more than 100 boxes for a single sample, only the top 100 will be considered.
- ⏱️ **Efficiency Tip**: Round all floating-point numbers in your submission to **two decimal places** to reduce file size and transmission overhead. (To ensure fairness during evaluation, all decimal numbers in the submitted predictions will be rounded to two decimal places.)
|