Ubuntu
commited on
Commit
β’
03fc49e
1
Parent(s):
23d56ee
fix README
Browse files
README.md
ADDED
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
annotations_creators:
|
3 |
+
- found
|
4 |
+
language_creators:
|
5 |
+
- found
|
6 |
+
languages:
|
7 |
+
- en
|
8 |
+
licenses:
|
9 |
+
- mit
|
10 |
+
multilinguality:
|
11 |
+
- monolingual
|
12 |
+
pretty_name: CLIP-Kinetics700
|
13 |
+
size_categories:
|
14 |
+
- 100K<n<1M
|
15 |
+
task_categories:
|
16 |
+
- feature-extraction
|
17 |
+
- zero-shot-classification
|
18 |
+
---
|
19 |
+
|
20 |
+
# Dataset Card for CLIP-Kinetics70
|
21 |
+
|
22 |
+
## Table of Contents
|
23 |
+
- [Table of Contents](#table-of-contents)
|
24 |
+
- [Dataset Description](#dataset-description)
|
25 |
+
- [Dataset Summary](#dataset-summary)
|
26 |
+
- [Dataset Preprocessing](#dataset-preprocessing)
|
27 |
+
- [Dataset Structure](#dataset-structure)
|
28 |
+
- [Data Instances](#data-instances)
|
29 |
+
- [Data Fields](#data-fields)
|
30 |
+
- [Data Splits](#data-splits)
|
31 |
+
- [Dataset Creation](#dataset-creation)
|
32 |
+
- [Source Data](#source-data)
|
33 |
+
- [Simple Experiments](#dataset-creation)
|
34 |
+
- [Zero-shot Evaluation](#zero-shot)
|
35 |
+
- [Linear-probe Evaluation](#zero-shot)
|
36 |
+
|
37 |
+
## Dataset Description
|
38 |
+
### Dataset Summary
|
39 |
+
|
40 |
+
CLIP-Kinetics700 is a compressed version of the Kinetics700 dataset using OpenAI's CLIP model.
|
41 |
+
|
42 |
+
The original dataset is ~700 GB making it difficult to use and hold in memory on one machine. By downsampling each video to 1 FPS and encoding the frames using CLIP we we're able to compress the dataset to ~8 GB making it very memory-friendly and easy to use.
|
43 |
+
|
44 |
+
### Dataset Preprocessing
|
45 |
+
|
46 |
+
[clip-video-encode](https://github.com/iejMac/clip-video-encode) is a tool you can use to easily and efficiently compute CLIP embeddings from video frames. We used it to generate the embeddings for this dataset.
|
47 |
+
|
48 |
+
## Dataset Structure
|
49 |
+
|
50 |
+
### Data Format
|
51 |
+
|
52 |
+
We formatted this as a [WebDataset](https://github.com/webdataset/webdataset) for better data-loading performance when training the models.
|
53 |
+
Each split contains a list of tar files each with 10000 data samples. This format can be read and used easily using the EmbeddingWebDatasetReader from [clip-video-encode](https://github.com/iejMac/clip-video-encode).
|
54 |
+
|
55 |
+
```
|
56 |
+
CLIP-Kinetics700
|
57 |
+
βββ splits.csv
|
58 |
+
βββ ds_00000.tar
|
59 |
+
| βββ vid_00000.npy
|
60 |
+
| βββ vid_00000.txt
|
61 |
+
| βββ vid_00000.json
|
62 |
+
| βββ vid_00001.npy
|
63 |
+
| βββ vid_00001.txt
|
64 |
+
| βββ vid_00001.json
|
65 |
+
| βββ ...
|
66 |
+
| βββ vid_10000.npy
|
67 |
+
| βββ vid_10000.txt
|
68 |
+
| βββ vid_10000.json
|
69 |
+
βββ ds_00001.tar
|
70 |
+
| βββ vid_10001.npy
|
71 |
+
| βββ vid_10001.txt
|
72 |
+
| βββ vid_10001.json
|
73 |
+
β ...
|
74 |
+
...
|
75 |
+
```
|
76 |
+
|
77 |
+
|
78 |
+
### Data Fields
|
79 |
+
* vid.npy: the numpy array with the per-frame embeddings. Shape -> (n_frames, 512)
|
80 |
+
* vid.cap: the "caption" of the video. In this case it is the Kinetics700 label.
|
81 |
+
* vid.json: additional metadata - YouTube video ID, start time, end time.
|
82 |
+
|
83 |
+
### Data Splits
|
84 |
+
* Train - 536489 samples | 54 tar's
|
85 |
+
* Validation - 33966 samples | 4 tar's
|
86 |
+
* Test - 64532 samples | 7 tar's
|
87 |
+
|
88 |
+
## Dataset Creation
|
89 |
+
|
90 |
+
### Source Data
|
91 |
+
|
92 |
+
Data was sourced from DeepMind's [Kinetics700](https://www.deepmind.com/open-source/kinetics) dataset and downloaded using [this](https://github.com/cvdfoundation/kinetics-dataset) convenient repository.
|
93 |
+
|
94 |
+
## Simple Experiments
|
95 |
+
|
96 |
+
Using [this repository](https://github.com/LAION-AI/temporal-embedding-aggregation) we evaluate CLIP-Kinetics700 with the following simple methods:
|
97 |
+
|
98 |
+
### [Zero-shot Evaluation](https://github.com/LAION-AI/temporal-embedding-aggregation/blob/master/src/evaluation/zero_shot.py)
|
99 |
+
|
100 |
+
| | Accuracy |
|
101 |
+
| ---------------- | -------- |
|
102 |
+
| Top-1 | 0.22 |
|
103 |
+
| Top-5 | 0.43 |
|
104 |
+
| mean(Top1, Top5) | 0.33 |
|
105 |
+
|
106 |
+
### [Linear-probe Evaluation](https://github.com/LAION-AI/temporal-embedding-aggregation/blob/master/src/evaluation/linear_probe.py)
|
107 |
+
|
108 |
+
| | Accuracy |
|
109 |
+
| ---------------- | -------- |
|
110 |
+
| Top-1 | 0.41 |
|
111 |
+
| Top-5 | 0.65 |
|
112 |
+
| mean(Top1, Top5) | 0.53 |
|