melodyyyyy123/PRS-Group28-data

PRS Group 28 · Indoor Dog Behaviour, Human–Dog Interaction and Dog Pose Data

Data used by the PRS Group 28 pet recognition system. The system has three models:

• Model One – dog/person detection and pose estimation (RTMDet + RTMPose, AP-10K keypoints for dogs, COCO keypoints for people). It turns every clip into a skeleton sequence.
• Model Two – dog behaviour recognition from the dog skeleton stream.
• Model Three – human–dog interaction recognition from both skeleton streams.

This repository holds the short video clips with their label manifests, the skeleton sequences Model One produced from them, extra skeletons from APT-36K videos, and the hand-labelled frames used to evaluate the dog pose model.

Contents

Folder	Files	What it is
data/dog_behavior/	202 MP4 + dog_behavior_clips.csv	Single-dog clips with one behaviour label each (~261 MB)
data/interaction/	147 MP4 + clips.csv	One person + one dog clips with one interaction label each (~350 MB)
data/skeletons_generated/	348 NPZ	Model One skeletons for the clips above (dog_<clip_id>.npz, interaction_<clip_id>.npz)
data/skeletons_generated_APT/	80 NPZ + apt_behavior_clips.csv	Model One skeletons for APT-36K dog clips, with behaviour labels; no video included
data/model one evaluation frame/	36 JPG + 36 JSON	Hand-labelled dog boxes and 17 keypoints (LabelMe format) for pose evaluation

data/
├── dog_behavior/
│   ├── dog_behavior_clips.csv
│   └── dXXX_vXX_cXXX.mp4
├── interaction/
│   ├── clips.csv
│   └── dXX_pXX_vXX_cXXX.mp4
├── skeletons_generated/
│   ├── dog_<clip_id>.npz
│   └── interaction_<clip_id>.npz
├── skeletons_generated_APT/
│   ├── apt_behavior_clips.csv
│   └── apt_<clip_id>.npz
└── model one evaluation frame/
    ├── dXXX_vXX_cXXX_fNNNN.jpg
    └── dXXX_vXX_cXXX_fNNNN.json

Every MP4 has exactly one manifest row and vice versa. Every clip has a skeleton file except d025_v24_c002 (night-time infrared, detector could not find the dog; marked valid = 0).

Sources

The clips were cut by hand from project-contributed recordings and from publicly available videos on YouTube, Bilibili, Pexels and Mixkit, and from the InterPet and SmartHomeBench research collections. The source of each clip is recorded in the notes column (for example source_bili_<BV id>, source_yt_<video id>, source_pexels_...). Indoor, third-person, fixed or stable camera footage was preferred; a few outdoor or partly occluded clips were kept where the action was unambiguous.

data/skeletons_generated_APT/ comes from dog videos in APT-36K. Only the skeletons Model One extracted and our behaviour labels are included, not the APT-36K frames or annotations.

People and dogs are identified only by anonymous ids (dXXX, pXX). No names, breeds or other identity information is stored.

1. Dog behaviour clips (data/dog_behavior/)

202 clips from 68 source videos and 66 dogs. Duration 1.3–10.3 s (mean 5.4 s, 18.2 min in total). Each clip shows one dog doing one stable behaviour; when posture and activity conflict, the posture label wins (a dog chewing a toy while lying down is lying_sleeping).

behavior_label	Meaning	Clips	Valid
walking_run	Walking or running, continuous gait	30	28
standing	Standing still on four legs	36	36
sitting	Hindquarters on the ground, front legs supporting	37	36
lying_sleeping	Trunk on the ground, awake or asleep	50	47
eating_drinking	Head at food or water with visible intake	23	23
playing_alone	Dog plays alone with a toy or object	26	26
Total		202	196

playing_alone could not be labelled consistently and is not used by the released Model Two, which has five classes.

dog_behavior_clips.csv columns:

Column	Example	Description
clip_id	d001_v01_c001	Clip id, same as the MP4 file name
source_video_id	d001_v01	Original video the clip was cut from
dog_id	d001	Anonymous dog id
start_sec, end_sec	163.661, 174.005	Position in the original video
duration_sec	10.344	end_sec − start_sec
behavior_label	walking_run	One of the six labels above
split	(empty)	Not filled here; see Splits
valid	1	1 usable, 0 excluded (multiple dogs, unusable skeleton, …)
notes	source_bili_BV1vNzaB2Ey2_P1	Source reference, then ;-separated quality notes

2. Human–dog interaction clips (data/interaction/)

147 clips from 82 source videos, 70 dogs and 71 people. Duration 1.2–10.2 s (mean 4.8 s, 11.8 min in total). Each clip has one main person and one main dog and one stable interaction label; label changes are cut into separate clips.

interaction	Meaning	Clips	Valid
no_interaction	Both present, but no sustained contact, hand feeding or shared play (talking, looking or gestures only)	33	33
petting	Hand stays on the dog's head or trunk with small stroking movements; holding or hugging included, brief touches excluded	39	39
playing_together	Both actively engaged: chasing, tug of war, repeated large movements; a dog playing alone does not count	40	38
hand_feeding	Food passed from the person's hand or a hand-held container to the dog; eating from a bowl does not count	35	34
Total		147	144

The training code calls hand_feeding feeding; map it when building a training manifest. The other three names are unchanged.

clips.csv columns: clip_id, source_video_id, dog_id, person_id, start_sec, end_sec, duration_sec, dog_behavior, interaction, split, valid, notes. They mean the same as above, plus:

• person_id – anonymous person id.
• dog_behavior – coarse auxiliary dog label (play, eat_drink, other). It is not a substitute for the behaviour dataset.

Interaction ids use two-digit numbers (d01_p01_v01_c001); they are independent of the three-digit ids in dog_behavior/, so d01 here and d001 there are not necessarily the same dog.

3. Skeleton sequences (data/skeletons_generated/, data/skeletons_generated_APT/)

Each clip was resampled to 15 fps and passed through Model One: RTMDet finds the dog (and, for interaction clips, the person), a tracking rule keeps one target per stream, and RTMPose estimates 17 keypoints inside each box. Each NPZ contains:

Array	Shape	Description
skeletons	(2, T, 17, 3) float32	Stream 0 = person (COCO keypoint order), stream 1 = dog (AP-10K order); x, y in pixels, keypoint confidence
boxes	(2, T, 5) float32	Person / dog box x1, y1, x2, y2, score
fps	scalar int	Always 15
image_size	(2,) int	Source width, height in pixels

T is the number of frames at 15 fps. A stream is all zeros in frames where its target was not found. dog_* files and all APT files have an all-zero person stream. Coordinates are in pixels, not normalised, and can fall outside the image for body parts out of frame.

apt_behavior_clips.csv has clip_id, behavior_label, valid and notes. Its 80 clips are mostly walking_run (67), plus standing (11) and lying_sleeping (2); 74 are valid. They were added to give Model Two more walking and standing examples. To train with them, put both skeleton folders in one directory and rename each APT file from apt_<id>.npz to dog_apt_<id>.npz.

4. Pose evaluation frames (data/model one evaluation frame/)

36 frames (1 dog each) from 28 dogs in 28 source videos, 6 frames per behaviour label. They were used to choose the dog pose weights for Model One. None of these frames or their source videos were used to train a pose model.

Each .json is a LabelMe 7.8 file with one rectangle labelled dog (the box) and up to 17 point shapes named with the AP-10K keypoints:

Index	Name	Index	Name	Index	Name
0	L_Eye	6	L_Elbow	12	L_Knee
1	R_Eye	7	L_F_Paw	13	L_B_Paw
2	Nose	8	R_Shoulder	14	R_Hip
3	Neck	9	R_Elbow	15	R_Knee
4	Root_of_tail	10	R_F_Paw	16	R_B_Paw
5	L_Shoulder	11	L_Hip

Left and right are the dog's own left and right. Only visible keypoints with a clear position were labelled; occluded or out-of-frame joints are simply absent (as in AP-10K) and should be treated as unlabelled. 547 of the 612 keypoint slots (89%) are labelled. Model predictions were used as a starting point and every point was corrected by hand by one annotator.

Naming

d = dog, p = person, v / t = source video, c = clip, f = frame index within the clip.

Object	Pattern	Example
Behaviour clip	dXXX_vXX_cXXX	d001_v01_c001.mp4
Interaction clip	dXX_pXX_vXX_cXXX	d01_p01_v01_c001.mp4
Skeleton	dog_<clip_id> / interaction_<clip_id> / apt_<clip_id>	dog_d001_v01_c001.npz
Evaluation frame	<clip_id>_fNNNN	d001_v01_c001_f0090.jpg

Splits

The split column is left empty in this repository. The frozen splits used by the released models are in the project code repository under pet_recognition_system/outputs/best_models/manifests/:

• interaction_manifest_mixed_v2.csv – 121 valid clips: 75 train / 22 val / 24 test (6 per class). No dog_id, person_id or source_video_id appears in more than one split.
• dog_behavior_manifest_5class_v4.csv – 201 valid clips over five classes: 137 development clips with a cv_fold column for five-fold cross-validation, 31 APT clips used for training only, and 33 test clips.

These manifests are stricter than the valid flags here (for example, they drop playing_alone), so their counts are lower.

If you make your own split, group by dog_id and source_video_id (and person_id for interactions), never by clip or frame: clips from one video share background and subject and leak across a random split. Freeze the split before any sampling, augmentation or skeleton extraction.

Loading example

import numpy as np, pandas as pd

clips = pd.read_csv("data/dog_behavior/dog_behavior_clips.csv")
clips = clips[clips.valid == 1]

z = np.load("data/skeletons_generated/dog_d001_v01_c001.npz")
dog = z["skeletons"][1]   # (T, 17, 3): x, y, confidence in AP-10K order
print(dog.shape, int(z["fps"]), z["image_size"])

Limitations

• Small: about 200 behaviour clips and 150 interaction clips, many dogs with only one or two clips.
• Classes are not perfectly balanced, and the APT clips are mostly walking.
• Skeletons are model output, not ground truth; tracking errors and low-confidence joints remain, especially for occlusion, night-vision and multi-dog scenes.
• The pose evaluation set has 36 frames and one annotator, and does not score occluded joints.
• Labels were made by the project team and have not been checked by a second annotator.

License and use

The annotations, manifests, skeletons and hand-labelled keypoints are released under CC BY 4.0. The video clips are excerpts of third-party videos and remain subject to the terms of their original sources (see notes); they are shared for non-commercial research and education only. APT-36K content is subject to the APT-36K license. If you are a rights holder and want a clip removed, please open a discussion on this repository.