RTMPose Custom Data Training

Train an RTMPose model on your own dataset with a custom keypoint schema. This workspace includes a complete example for 18 keypoints (ALF pose data), from raw YOLO labels through training and inference.

Annotate your data: Use Keypoints Studio — a desktop GUI for building YOLO pose datasets (custom keypoint schemas, bbox + keypoint editing, YOLO-pose .txt export).

Overview

Item	Value
Model	RTMPose-M (CSPNeXt backbone)
Keypoints	18 (custom schema)
Input size	256×192 (default) or 448×448
Epochs	200 (stage-2 pipeline from epoch 180)
Data format	COCO JSON (top-down, single person per image)

Pipeline summary

Raw images + YOLO labels  →  COCO JSON + folder layout  →  Train RTMPose  →  Inference

---

Project Structure

mmpose/
├── DataPreprocessing/
│   ├── custom18_keypoints.py      # Keypoint names, skeleton, count
│   └── prepare_dataset.py         # YOLO → COCO conversion script
├── configs/
│   ├── _base_/datasets/custom18.py           # MMPose dataset metainfo
│   └── body_2d_keypoint/rtmpose/custom/
│       ├── rtmpose-m_8xb64-200e_custom18-256x192.py
│       └── rtmpose-m_8xb64-200e_custom18-448x448.py
├── data/custom_pose/              # Prepared dataset (created by script)
├── tools/train.py                 # Training entry point
├── demo/                          # Inference demos
└── work_dirs/                     # Checkpoints and logs

---

Step 1 — Environment Setup

Requirements

• Windows 10/11 (64-bit)
• NVIDIA GPU + driver (CUDA recommended)
• Python 3.10 via Conda

Install

conda create -n rtmpose python=3.10 -y
conda activate rtmpose
python -m pip install -U pip

# PyTorch (CUDA 11.8)
pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu118

# Core dependencies (pinned for stability)
pip install "numpy==1.26.4" "setuptools<81" "opencv-python==4.11.0.86"
pip install -U openmim
pip install mmengine==0.10.7 mmcv==2.1.0 mmdet==3.3.0
pip install xtcocotools==1.14.3 --no-build-isolation --no-cache-dir
pip install albumentations==1.4.24 charset-normalizer

# Install this repo
cd path\to\mmpose
pip install -e .

Note: Keep numpy<2. Upgrading NumPy to 2.x breaks xtcocotools and other compiled extensions.

---

Step 2 — Define Your Keypoints

Edit DataPreprocessing/custom18_keypoints.py to match your annotation schema.

Field	Purpose
NUM_KEYPOINTS	Total keypoint count
KEYPOINT_NAMES	Ordered list of keypoint names (index = annotation order)
SKELETON_LINKS	Pairs of keypoint indices for visualization and metainfo

Example (18 keypoints):

right_hand, left_hand, right_elbow, left_elbow,
right_shoulder, left_shoulder, Trachea, nose, forehead,
right_hip, left_hip, sacrum,
right_knee, left_knee, right_ankle, left_ankle,
right_footindex, left_footindex

The same definition must be reflected in configs/_base_/datasets/custom18.py (MMPose metainfo: names, colors, skeleton, swap pairs for horizontal flip).

---

Step 3 — Prepare Raw Data

Annotate with Keypoints Studio (recommended)

Label images and export YOLO-pose .txt files using Keypoints Studio:

• Custom keypoint schemas (add / remove / reorder points)
• Bounding box + keypoint editing with preview
• Export YOLO-pose labels alongside images

Place the exported folders under your source directory (default: alf_data/).

Input format (YOLO pose)

Place images and labels under a source folder (default: alf_data/):

alf_data/
└── <sequence_name>/
    ├── image001.jpg
    ├── image001.txt
    └── ...

Each .txt line is one person:

class  x_center  y_center  width  height  (kpt_x  kpt_y  visibility) × N

• Coordinates are normalized (0–1) relative to image size.
• visibility: 0 = not labeled, 1 = occluded, 2 = visible.
• Detection-only lines (5 values, no keypoints) are skipped.

Convert to MMPose format

python DataPreprocessing/prepare_dataset.py --src alf_data --out data/custom_pose --val-ratio 0.1

Output:

data/custom_pose/
├── images/train/
├── images/val/
├── labels/train/
├── labels/val/
├── annotations/custom_train.json
├── annotations/custom_val.json
└── dataset_stats.json

Use --dry-run to scan and print statistics without writing files.

---

Step 4 — Dataset Metainfo

MMPose reads keypoint metadata from configs/_base_/datasets/custom18.py. This file defines:

• keypoint_info — name, id, color, left/right swap partner
• skeleton_info — bone connections for visualization
• joint_weights and sigmas — used in loss and COCO evaluation

When you change the number or order of keypoints, update both:

1. DataPreprocessing/custom18_keypoints.py
2. configs/_base_/datasets/custom18.py

And set num_keypoints / out_channels in the training config.

---

Step 5 — Training Configuration

Two configs are provided:

Config	Input size (W×H)	Use case
rtmpose-m_8xb64-200e_custom18-256x192.py	192 × 256	Default, faster, less VRAM
rtmpose-m_8xb64-200e_custom18-448x448.py	448 × 448	Higher resolution, more VRAM

Key settings in the config:

Setting	Default	Description
max_epochs	200	Total training epochs
stage2_num_epochs	20	Last N epochs use a lighter augmentation pipeline
val_interval	10	Validate every N epochs
data_root	data/custom_pose/	Prepared dataset path
num_keypoints	18	Must match your schema

The backbone loads a COCO-pretrained CSPNeXt-M checkpoint. Only the pose head is trained for your 18 keypoints.

---

Step 6 — Train

Download pretrained backbone (recommended)

python -c "import os,urllib.request; u='https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/cspnext-m_udp-aic-coco_210e-256x192-f2f7d6f6_20230130.pth'; os.makedirs('checkpoints',exist_ok=True); urllib.request.urlretrieve(u, 'checkpoints/'+os.path.basename(u))"

Start training

Use the environment Python directly (more reliable than mim on Windows):

cd path\to\mmpose

python tools/train.py configs/body_2d_keypoint/rtmpose/custom/rtmpose-m_8xb64-200e_custom18-256x192.py ^
  --work-dir work_dirs/custom18_rtmpose-m ^
  --cfg-options ^
    model.backbone.init_cfg.checkpoint="checkpoints/cspnext-m_udp-aic-coco_210e-256x192-f2f7d6f6_20230130.pth" ^
    train_dataloader.batch_size=8 ^
    val_dataloader.batch_size=8 ^
    train_dataloader.num_workers=0 ^
    val_dataloader.num_workers=0 ^
    train_dataloader.persistent_workers=False ^
    val_dataloader.persistent_workers=False

Batch size guide (Quadro P1000, 4 GB VRAM):

Input size	Suggested batch size
256×192	8 (try 4 if OOM)
448×448	2 (try 4 if stable)

Checkpoints are saved to work_dirs/custom18_rtmpose-m/. The best model (by validation AP) is kept automatically.

Monitor training

Logs appear in the terminal and in work_dirs/<run>/<timestamp>.log. Example:

Epoch(train) [1][50/2371]  loss: 0.090868  acc_pose: 0.037325

Stop training

Press Ctrl+C in the terminal, or:

tasklist | findstr /i python
taskkill /PID <pid> /F

---

Step 7 — Inference with Your Model

After training, use the saved checkpoint with your custom config:

python demo/topdown_demo_with_mmdet.py ^
  projects/rtmpose/rtmdet/person/rtmdet_nano_320-8xb32_coco-person.py ^
  https://download.openmmlab.com/mmpose/v1/projects/rtmpose/rtmdet_nano_8xb32-100e_coco-obj365-person-05d8511e.pth ^
  configs/body_2d_keypoint/rtmpose/custom/rtmpose-m_8xb64-200e_custom18-256x192.py ^
  work_dirs/custom18_rtmpose-m/best_coco_AP_epoch_*.pth ^
  --input your_video.mp4 ^
  --output-root vis_results ^
  --device cuda:0

Replace best_coco_AP_epoch_*.pth with the actual checkpoint filename from work_dirs/.

---

Adapting to a Different Dataset

To train on a new keypoint schema (not 18 ALF keypoints):

1. Rename or copy custom18_keypoints.py and set your keypoint names and skeleton.
2. Create a new metainfo file under configs/_base_/datasets/.
3. Copy a training config and update:
   • num_keypoints
   • model.head.out_channels
   • metainfo=dict(from_file='configs/_base_/datasets/your_dataset.py')
   • data_root and annotation paths
4. Write or adapt a conversion script if your labels are not YOLO pose format.
5. Train with the new config.

---

Troubleshooting

Error	Fix
albumentations is not installed	pip install albumentations==1.4.24
numpy.dtype size changed	pip install numpy==1.26.4 --force-reinstall --no-deps then reinstall xtcocotools
persistent_workers needs num_workers > 0	Add train_dataloader.persistent_workers=False (and val/test)
can't open file tools/train.py	pip install -e . from repo root
CUDA out of memory	Lower train_dataloader.batch_size (e.g. 4 or 2)
Training very slow on Windows	Use num_workers=0; first epoch is always slowest

---

References

• Keypoints Studio — keypoint annotation tool (YOLO pose export)
• RTMPose project docs
• MMPose custom dataset tutorial
• RTMPose paper / model zoo