Skip to content

Commit

Permalink
Update convert_datasets.md
Browse files Browse the repository at this point in the history
  • Loading branch information
Wei-Chen-hub authored Aug 23, 2023
1 parent 3e24142 commit bd89546
Showing 1 changed file with 78 additions and 5 deletions.
83 changes: 78 additions & 5 deletions docs/convert_datasets.md
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
# Convert Datasets to HumanData and Use in Training
# Convert Datasets to HumanData and Preparation Before Training
**This page describes how to use the various HumanData flies** as well as (only if you are interested) how to process various datasets to current HumanData.

## Introduction & Settings
Expand All @@ -23,7 +23,7 @@
git clone -b convertors https://github.com/open-mmlab/mmhuman3d.git
cd /Your_path_to_mmhuman3d/mmhuman3d # e.g. /mnt/d/zoehuman/mmhuman3d
```
6. (For convert only) The usage of convert_datasets.py (Example)
6. (For convert only) The usage of convert_datasets.py (Example), if "--modes" is not specified, all avalilable modes (all data) will be processed.
```
python tools/convert_datasets.py \
--datasets <dataset name/perfix> \
Expand All @@ -32,9 +32,11 @@
--modes train
# dataset folder should be os.path.join(root_path, datasets)
# output path is the folder which sotres output HumanData
```
```
7. (For convert only) For discussing processing speed or time, hardware config is RTX3090 + i9-12900 + 64GB RAM + Cuda11.8 & Python3.9. The most resource-consuming processes are calling SMPL/SMPLX models and camera projection.

## Overview - Current Supported Datasets
- EgoBody
- GTA-Human++
- H36M (Neural Annot)
- MPII (Neural Annot)
Expand Down Expand Up @@ -66,6 +68,9 @@ D:\datasets\h36m\
```

**Step 2 (Converter) - Convert Dataset**

Due to size considerations, "train" and "val" set is separated in to 5 and 2 files (depends on subject id). HumanData file ranges from 900MB to 1.5GB. Processing speed is ~40item/s, total time is about **12 hours**.

```
python tools/convert_datasets.py \
--datasets h36m\
Expand Down Expand Up @@ -101,6 +106,8 @@ python tools/preprocess/neural_annot.py --dataset_path /YOUR_PATH/mpii
```

**Step 3 (Converter) - Convert Dataset**

Process speed is ~60item/s, totally ~17k item, and takes several minutes.
```
python tools/convert_datasets.py \
--datasets mpii \
Expand Down Expand Up @@ -138,7 +145,9 @@ This process converts the coco annotation json to faciliate sorting ids.
python tools/preprocess/neural_annot.py --dataset_path /YOUR_PATH/mscoco
```

**Step 3 (Converter) - Convert Dataset**
**Step 3 (Converter) - Convert Dataset**/

Processing speed is ~50item/s, total ~260k item.
```
python tools/convert_datasets.py \
--datasets mscoco \
Expand Down Expand Up @@ -178,6 +187,8 @@ python tools/preprocess/neural_annot.py --dataset_path /YOUR_PATH/pw3d
```

**Step 3 (Converter) - Convert Dataset**

3DPW datasets contains 3 modes ("train", "test" and "val"), processing speed is ~40item/s, total 60k item.
```
python tools/convert_datasets.py \
--datasets pw3d \
Expand All @@ -190,6 +201,68 @@ python tools/convert_datasets.py \
## Subsection 2 - Synthetic Datasets

<details>
<summary>GTA-Human++</summary>
<summary>AGORA (IP)</summary>

</details>


<details>
<summary>GTA-Human++ (IP)</summary>

</details>

## Subsection 3 - Real Multi-Human Datasets
<details>
<summary>EgoBody</summary>

**Step 1 - Only Step for using HumanData**

For Egobody dataset, please download as instructed in [HomePage](https://github.com/sanweiliti/EgoBody).
Egobody dataset can be split as two distinct subset, "egocentric" which denotes the views from human-wear kinect (single view, single person) and "kinect" which denotes the the fixed various kinect (multi-view, two person).

**Step 2 (Converter) - Convert Dataset**

There are 6 HumanData set, seperated by "train, test and val", and "egocentric and kinect". For example, "egocentric_train" & "kinect_train" comprise the train set as mentioned in the paper.

```
available_modes=['egocentric_train', 'egocentric_test', 'egocentric_val',
'kinect_train', 'kinect_test', 'kinect_val']
```

Kinect set conducts sequence-level processing whereas egocentric set only processes image-level and is much slower (due to changing camera parameters). Total processing time should be **1-3 days**.

```
python tools/convert_datasets.py \
--datasets egobody \
--root_path /mnt/d/datasets \
--output_path /mnt/d/datasets/egobody/output \
# --modes egocentric_train
```
</details>

<details>
<summary>Ubody (IP)</summary>

</details>



## Subsection 4 - Real Single-Human Datasets
<details>
<summary>SSP3D</summary>

**Step 1 - Only Step for using HumanData**

Download dataset from [HomePage](https://github.com/akashsengupta1997/SSP-3D)

**Step 2 (Converter) - Convert Dataset**

This is a very small dataset which should finish in a gilmpse.
```
python tools/convert_datasets.py \
--datasets ssp3d \
--root_path /mnt/e \
--output_path /mnt/e/ssp3d/output
```
</details>

0 comments on commit bd89546

Please sign in to comment.