tsinghua-ee/SAVEBench

---
license: apache-2.0
---
Audio:
1. LibriSpeech test clean full set
2. Audio caps test full set

Video:
1. NExTQA: nextqa_test.json
ID provided in the "image" field

Image:
1. Flickr30k: flickr30k_captions.json
(this is the standard 1k test set). ID provided in the "image" field.
2. TextVQA: textvqa.json
ID provided in the "image" field
3. GQA: testdev_balanced_questions_with_images.json
ID provided in the "image" field

Audio-visual:
1. How2: how2_test.json
ID provided in "image". Format: <video_id>_<start_second>_<end_second>.mp4 or .wav.
2. Audio-Visual Sound Source Detection (AVSSD): testdata_formatted.json
ID provided in the "image" field. The first one is image and the second one is the corresponding audio.
3. Audio Visual Matching (AVM): audiovisualmatching_combined.json
ID provided in the "image" field as a list of two values. The first one is the image and the second one is the audio/speech
Whether it is from VGGSS or is from SpokenCOCO is indicated in the ID as well
4. Audio-visual question answering (AVQA) Ego4D-QA: ego4d_qa.json
Video ID indicates the frame index
5. Audio-visual question answering (AVQA) Presentation-QA: presentation_qa.json

许可证：Apache 2.0

音频：
1. LibriSpeech 清洁测试全集
2. 音频字幕(Audio Caps)测试全集

视频：
1. NExTQA：nextqa_test.json，其ID存储于"image"字段中。

图像：
1. Flickr30k：flickr30k_captions.json，此为标准1k测试集，其ID存储于"image"字段中。
2. TextVQA：textvqa.json，其ID存储于"image"字段中。
3. GQA：testdev_balanced_questions_with_images.json，其ID存储于"image"字段中。

音视频：
1. How2：how2_test.json，其ID存储于"image"字段中，文件命名格式为<video_id>_<start_second>_<end_second>.mp4或.wav。
2. 音视频声源检测(Audio-Visual Sound Source Detection, AVSSD)：testdata_formatted.json，其ID存储于"image"字段中，列表第一项为对应图像，第二项为对应音频。
3. 音视频匹配(Audio Visual Matching, AVM)：audiovisualmatching_combined.json，其ID以双值列表的形式存储于"image"字段中，第一项为图像，第二项为音频/语音；样本来源为VGGSS或SpokenCOCO，具体信息可通过ID识别。
4. 音视频问答(Audio-visual question answering, AVQA) Ego4D-QA：ego4d_qa.json，视频ID对应帧索引。
5. 音视频问答(Audio-visual question answering, AVQA) Presentation-QA：presentation_qa.json