File size: 6,466 Bytes
4ceeb65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
---
pretty_name: EventDrive
language:
- en
---

# EventDrive

## EventDrive: Event Cameras for Vision-Language Driving Intelligence

Dongyue Lu, Rong Li, Ao Liang, Lingdong Kong, Wei Yin, Lai Xing Ng, Benoit R. Cottereau, Camille Simon Chane, and Wei Tsang Ooi

**CVPR 2026**

[Project Page](https://dylanorange.github.io/projects/eventdrive/) | [Paper](https://dylanorange.github.io/projects/eventdrive/static/files/EventDrive.pdf) | [Dataset](https://huggingface.co/datasets/dylanorange/EventDrive)

EventDrive is a unified event-frame driving benchmark for vision-language driving intelligence. It combines synchronized RGB frames, event-camera data, and instruction-style annotations to study how event sensing supports multimodal perception, reasoning, prediction, and planning under diverse driving conditions.

The benchmark covers four dimensions:

- **Perception**: scene-level driving perception questions.
- **Understanding**: object awareness, grounding, appearance, status, and spatial-relation questions.
- **Prediction**: short-term behavior prediction for a highlighted dynamic agent.
- **Planning**: high-level driving intent and ego-trajectory prediction.

## Repository Layout

```text
.
├── eventdrive_perception.tar.gz
├── eventdrive_understanding.tar.gz
├── eventdrive_prediction.tar.gz
├── eventdrive_planning.tar.gz
├── json/
│   ├── perception/
│   │   ├── dsec/
│   │   ├── m3ed/
│   │   └── pku/
│   ├── understanding/
│   ├── prediction/
│   └── planning/
└── scripts/
    ├── evaluation_perception.py
    ├── evaluation_understanding.py
    ├── evaluation_prediction.py
    └── evaluation_planning.py
```

Create a `data/` directory and extract all archives from the repository root:

```bash
mkdir -p data
tar -xzf eventdrive_perception.tar.gz -C data
tar -xzf eventdrive_understanding.tar.gz -C data
tar -xzf eventdrive_prediction.tar.gz -C data
tar -xzf eventdrive_planning.tar.gz -C data
```

The extracted data follows this structure:

```text
data/
├── perception/
│   ├── dsec/
│   │   ├── train/<sequence>/{image,event}/
│   │   └── test/<sequence>/{image,event}/
│   ├── m3ed/<sequence>/{image,event}/
│   └── pku/aps_frames_sampled/val/<condition>/<sequence>/
├── understanding/
│   ├── train/<sequence>/{image,event}/
│   └── test/<sequence>/{image,event}/
├── prediction/<sequence>/{image,event}/
└── planning/<sequence>/{image,event}/
```

Each `image/` directory contains RGB frames. Each `event/` directory contains the paired event-camera representation in `.npz` format.

For PKU perception data, paired `.png` and `.npz` files are stored side by side in each sequence directory instead of separate `image/` and `event/` directories.

## Annotation Files

All annotation paths are relative to the repository root and start with `data/`. Run the scripts from the repository root after extracting the archives.

Each dimension provides train and test annotations. Files ending in `_hard.json` contain the hard test subsets.

The released annotations under `json/` use separate prompts for answer components such as option letter and label text, or speed and path intent. Samples originating from the same question share an `original_id`. The evaluation scripts use this field to pair component predictions before computing joint accuracy. Planning trajectory samples are evaluated independently and do not require an `original_id`.

A typical annotation includes paired image and event paths plus an instruction-answer conversation:

```json
{
  "image": "data/perception/dsec/test/interlaken_00_a/image/000005.png",
  "event": "data/perception/dsec/test/interlaken_00_a/event/000005.npz",
  "category": "Scene type",
  "original_id": "perception/dsec/dsec_test_perception.json:000000",
  "subtask": "option_letter",
  "conversations": [
    {
      "from": "human",
      "value": "<instruction>"
    },
    {
      "from": "gpt",
      "value": "<ground-truth answer>"
    }
  ]
}
```

Add a `model_output` field to each sample after inference:

```json
{
  "model_output": "<model prediction>"
}
```

For understanding grounding samples, boxes use `[x, y, w, h]`, where `(x, y)` is the top-left corner and `(w, h)` is the width and height.

## Evaluation

Install the evaluation dependencies:

```bash
pip install numpy tqdm
```

Run the matching evaluator on an inference result JSON file generated from the annotations under `json/`:

```bash
python scripts/evaluation_perception.py \
  --pred-json results/dsec_test_perception.json

python scripts/evaluation_understanding.py \
  --pred-json results/dsec_test_understanding.json \
  --iou-thresh 0.6

python scripts/evaluation_prediction.py \
  --pred-json results/m3ed_test_prediction.json

python scripts/evaluation_planning.py \
  --pred-json results/m3ed_test_planning.json
```

The evaluators write summary JSON files next to the prediction file. They also save mismatch examples for debugging when applicable.

Metrics:

- **Perception**: joint accuracy after pairing the split option-letter and label-text answers. Both answers must be correct.
- **Understanding**: joint QA accuracy after pairing the split option-letter and label-text answers, category-wise accuracy, grounding accuracy at the selected IoU threshold, and mean IoU. Both QA answers must be correct. The default IoU threshold is `0.6`.
- **Prediction**: speed accuracy, path accuracy, class-wise accuracy, and joint speed-path accuracy after pairing split answers.
- **Planning**: high-level speed accuracy, path accuracy, class-wise accuracy, joint speed-path accuracy after pairing split answers, and trajectory L2 error at `1s`, `3s`, and `5s`.

Planning trajectory predictions must contain exactly 10 `[x, y]` waypoints at 0.5-second intervals. Evaluation terminates with an error if a trajectory prediction does not follow this format.

## Citation

```bibtex
@InProceedings{Lu_2026_CVPR,
  author    = {Lu, Dongyue and Li, Rong and Liang, Ao and Kong, Lingdong and Yin, Wei and Ng, Lai Xing and Cottereau, Benoit R. and Chane, Camille Simon and Ooi, Wei Tsang},
  title     = {EventDrive: Event Cameras for Vision-Language Driving Intelligence},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
}
```