ADODAS2026 Baseline

Official baseline implementation for the ADODAS grand challenge (ACMMM 2026).

Tasks

• Track A1: three binary labels for Depression / Anxiety / Stress.
• Track A2: 21 ordinal item predictions with scores in {0, 1, 2, 3}.

Environment

Create a conda environment from envs/adodas.yaml:

conda env create -f envs/adodas.yaml
conda activate adodas

Feature directory

The config field feature_root should point to the anonymized feature tree. The loader expects:

<feature_root>/<split>/<anon_school>/<anon_class>/<anon_pid>/
|- audio/
|  |- mel_mfcc/<session>/
|  |- vad/<session>/
|  |- ssl_embed/<audio_ssl_model_tag>/<session>/
|  `- egemaps/<session>/
`- video/
     |- headpose_geom/<session>/
     |- face_behavior/<session>/
     |- qc_stats/<session>/
     |- vad_agg/<session>/
     |- body_pose/<session>/
     |- global_motion/<session>/
     `- vision_ssl_embed/<video_ssl_model_tag>/<session>/

Training

python train.py --task <a1|a2> --config <config_yaml>

Inference

python infer.py --task <a1|a2> --checkpoint <path_to_best.pt> [--config <config_yaml>] [--split <split_name>] [--output <csv_path>]

Typical workflow:

1. Train a model and get best.pt under <output_dir>/runs/<run_name>/checkpoints/.
2. Run inference using that checkpoint.

Output Structure

For each run, directories are organized as:

<output_dir>/runs/<run_name>/
|- logs/
|- checkpoints/
|- calibration/
`- submissions/   (created only when submissions are written)

Annotations：

File Types

• A01： "The North Wind and the Sun" standardized reading passage.
  • 有一回，北风跟太阳在那儿争论谁的本领大。说着说着，来了一个过路的，身上穿了一件厚袍子。他们俩就商量好了，说谁能先叫这个过路的把他的袍子脱下来，就算是他的本领大。北风就使劲吹起来，拼命地吹。可是，他吹得越厉害，那个人就把他的袍子裹得越紧。到末了儿，北风没辙了，只好就算了。一会儿，太阳出来一晒，那个人马上就把袍子脱了下来。所以，北风不得不承认，还是太阳比他的本领大。
• B01：Please describe how your day went yesterday.
  • 请描述一下，你昨天过的怎么样?
• B02：Please describe your happiest memory from the past week.
  • 请描述一下，现在回想最近一周最开心的记忆?
• B03：Please describe your saddest memory from the past week.
  • 请描述一下，现在回想最近一周最悲伤的记忆?

Auxiliary Attributes

1. 家庭结构（Family structure）
   1. 1=核心家庭，Nuclear
   2. 2=大家庭, Extended
   3. 3=单亲家庭, Single-parent
   4. 4=重组家庭，Blended
   5. 5=隔代家庭，Skipped-generation
   6. 6=其他，Other
2. 是否是独生子女（Only child status）
   1. Whether the respondent is an only child (1：Yes/0：No).
3. 如非独生子女,是否感受到父母有所偏爱？（Parental favoritism If not an only child）
   1. 1=偏爱兄弟姐妹，Favoring siblings
   2. 2=无偏爱，No favoritism
   3. 3=偏爱自己，Favoring self
4. 本学期相比上个学期，学习成绩变动情况（Academic performance change Compared with previous semester）:
   1. 1=进步，Improved,
   2. 2=退步，Declined,
   3. 3=稳定，Stable.
5. 本学期相比上个学期，情绪变动情况（Emotional state change Compared with previous semester）:
   1. 1=变好，Better
   2. 2=变差，Worse
   3. 3=无变化，No change

Features Description

1. Audio

1.1 mel_mfcc

• Contents:
  • Log-Mel spectrogram: 80 dims
  • MFCC: 13 dims
• Time resolution: 25 Hz
• sequence.npz:
  • mel_features: (T, 80)
  • mfcc_features: (T, 13)
  • timestamps_ms: (T,)
  • valid_mask: (T,)
• Field meanings:
  • mel_features[:, 0:80]: log energy over 80 Mel filter banks
  • mfcc_features[:, 0:13]: 13 MFCC coefficients
• pooled statistics:
  • mean/std/p10/p50/p90 are computed separately for Mel and MFCC features
  • Theoretical total dimensionality: 80*5 + 13*5 = 465
  • Column naming patterns:
    • mel_00_mean ... mel_79_p90
    • mfcc_00_mean ... mfcc_12_p90

1.2 vad

• Extraction method: webrtcvad
• Time resolution: 25 Hz
• sequence.npz:
  • features: (T, 1)
• Single-dimension semantics:
  • Dimension 0: vad_decision
    • 1 indicates speech
    • 0 indicates silence / non-speech
• pooled.json fields:
  • speech_ratio
  • total_speech_duration
  • total_silence_duration
  • num_speech_segments
  • num_silence_segments
  • pause_count
  • mean_pause_duration
  • max_pause_duration
  • long_pause_count
  • Other outputs:
    • segments.json: start/end timestamps of speech and silence segments

1.3 egemaps

• Extraction method: openSMILE eGeMAPSv02 functionals
• Total pooled dimensionality: 88
• Description:
  • These 88 dimensions are standardized acoustic statistical features covering F0, loudness, spectral slope, spectral flux, formant-related statistics, and voiced/unvoiced segment statistics.
• Example column names:
  • F0semitoneFrom27.5Hz_sma3nz_amean
  • F0semitoneFrom27.5Hz_sma3nz_stddevNorm
  • loudnessPeaksPerSec
  • VoicedSegmentsPerSec
  • MeanVoicedSegmentLengthSec
  • equivalentSoundLevel_dBp

1.4 ssl_embed

• Extraction method: final-layer hidden states from a speech self-supervised model, linearly interpolated and resampled to 25 Hz
• sequence.npz:
  • features: (T, D)
  • timestamps_ms: (T,)
  • valid_mask: (T,)
  • embed_dim: scalar
  • model_name: scalar
• Per-frame semantics:
  • Each frame is a D-dimensional speech representation vector produced by the corresponding pretrained model.
• pooled statistics:
  • mean/std/p10/p50/p90 are computed for each embedding dimension
  • Total dimensionality: 5 * D
  • Column naming pattern: embed_0000_mean ... embed_(D-1)_p90

Supported models and dimensions:

model_tag	HuggingFace / model identifier	Per-frame dimension
wavlm-base	microsoft/wavlm-base	768
chinese-hubert-base	TencentGameMate/chinese-hubert-base	768
chinese-hubert-large	TencentGameMate/chinese-hubert-large	1024
chinese-wav2vec2-base	TencentGameMate/chinese-wav2vec2-base	768
chinese-wav2vec2-large	TencentGameMate/chinese-wav2vec2-large	1024
wav2vec2-chinese-xlsr	jonatasgrosman/wav2vec2-large-xlsr-53-chinese-zh-cn	1024

2. Video Features

2.1 qc_stats

• Pretrained dependency: indirectly depends on upstream InsightFace buffalo_l outputs (face_meta)
• Extraction method: quality statistics computed from face_meta.parquet
• sequence.npz:
  • features: (T, 4)
  • feature_names = [quality_score, blur_score, brightness, det_score]
• 4-D mapping:
  • Dimension 0: quality_score, overall quality score
  • Dimension 1: blur_score, blur/sharpness score
  • Dimension 2: brightness
  • Dimension 3: det_score, face detection confidence
• Main pooled.json fields:
  • total_frames
  • detected_frames
  • detection_rate
  • valid_frames
  • valid_ratio
  • is_low_quality
  • blur_mean, blur_std, blur_min, blur_max
  • brightness_mean, brightness_std
  • quality_mean, quality_std
  • det_score_mean, det_score_std
  • face_area_ratio_mean
  • num_selected_frames

2.2 headpose_geom

• Pretrained dependency: indirectly depends on 5-point landmarks from upstream InsightFace buffalo_l
• Extraction method: geometric approximation based on 5-point landmarks
• sequence.npz:
  • features: (T, 5)
  • feature_names = [yaw, pitch, roll, ear_mean, mar]
• 5-D mapping:
  • Dimension 0: yaw, left-right head rotation
  • Dimension 1: pitch, up-down head motion
  • Dimension 2: roll, head tilt
  • Dimension 3: ear_mean, mean EAR of both eyes, approximately reflecting eye openness
  • Dimension 4: mar, mouth opening ratio
• pooled statistics:
  • mean/std/min/max are computed for each of yaw/pitch/roll/ear_mean/mar
  • Plus valid_ratio
  • Total dimensionality: 5*4 + 1 = 21

2.3 face_behavior

• Pretrained dependency: indirectly depends on upstream InsightFace buffalo_l outputs (face_meta)
• Extraction method: behavior statistics based on ear / mar / yaw / pitch / quality_score
• sequence.npz:
  • features: (T, 5)
• 5-D mapping:
  • Dimension 0: ear
  • Dimension 1: mar
  • Dimension 2: yaw
  • Dimension 3: pitch
  • Dimension 4: quality_score
• pooled.parquet statistical fields:
  • blink_count
  • blink_rate_per_min
  • avg_blink_duration_frames
  • mouth_open_ratio
  • mouth_movement_std
  • speech_activity_ratio
  • gaze_stability_score
  • yaw_range
  • pitch_range
  • saccade_count
  • expression_variability
  • expression_change_count
• Total dimensionality: 12

2.4 vad_agg

• Pretrained model: no
• Extraction method: align and aggregate audio vad onto the video timeline
• sequence.npz:
  • features: (T, 4)
  • feature_names = [speech_prob, speech_activity, local_speech_ratio, speech_transition]
• 4-D mapping:
  • Dimension 0: speech_prob, aligned speech probability/intensity
  • Dimension 1: speech_activity, binary speech activity
  • Dimension 2: local_speech_ratio, local-window speech ratio
  • Dimension 3: speech_transition, frame-to-frame speech state transition intensity
• pooled.parquet statistical fields:
  • speech_ratio
  • avg_speech_prob
  • speech_transitions
  • silence_ratio
  • local_speech_ratio_mean
  • local_speech_ratio_std

2.5 body_pose

• Pretrained model: MediaPipe PoseLandmarker
  • Preferred: pose_landmarker_full.task
  • Fallback: pose_landmarker_lite.task
• sequence.npz:
  • features: (T, 27)
  • landmark_names = [nose, left_shoulder, right_shoulder, left_elbow, right_elbow, left_wrist, right_wrist, left_hip, right_hip]
• 27-D mapping:
  • Each landmark outputs 3 dimensions: (x, y, visibility)
  • There are 9 upper-body landmarks, so total dimensionality is 27 = 9 * 3
• Flattened order:
  • nose_x, nose_y, nose_visibility
  • left_shoulder_x, left_shoulder_y, left_shoulder_visibility
  • right_shoulder_x, right_shoulder_y, right_shoulder_visibility
  • left_elbow_x, left_elbow_y, left_elbow_visibility
  • right_elbow_x, right_elbow_y, right_elbow_visibility
  • left_wrist_x, left_wrist_y, left_wrist_visibility
  • right_wrist_x, right_wrist_y, right_wrist_visibility
  • left_hip_x, left_hip_y, left_hip_visibility
  • right_hip_x, right_hip_y, right_hip_visibility
• Main pooled.json fields:
  • num_frames
  • num_valid_frames
  • valid_ratio
  • landmarks_mean: 27 dimensions
  • landmarks_std: 27 dimensions

2.6 global_motion

• Pretrained model: no
• Extraction method: OpenCV Farneback optical flow + frame differencing
• sequence.npz:
  • features: (T, 4)
  • feature_names = [flow_mag_mean, flow_mag_std, flow_angle_mean, frame_diff_mean]
• 4-D mapping:
  • Dimension 0: flow_mag_mean, mean optical-flow magnitude
  • Dimension 1: flow_mag_std, std of optical-flow magnitude
  • Dimension 2: flow_angle_mean, mean optical-flow direction
  • Dimension 3: frame_diff_mean, mean pixel difference between adjacent frames
• Main pooled.json fields:
  • num_frames
  • flow_magnitude_mean
  • flow_magnitude_std
  • flow_magnitude_max
  • frame_diff_mean
  • frame_diff_max
  • motion_energy

2.7 vision_ssl_embed

• Pretrained model: yes
• Extraction method: aligned face images are fed into a pretrained vision model; pooler_output or CLS token is used
• sequence.npz:
  • features: (T, D)
  • timestamps_ms: (T,)
  • valid_mask: (T,)
  • embed_dim: scalar
  • model_name: scalar
• Per-frame semantics:
  • Each frame is a D-dimensional high-level visual representation vector encoding face appearance, pose, texture, and expression-related patterns. It does not correspond to a single handcrafted geometric variable.
• pooled statistics:
  • mean/std/p10/p50/p90 are computed for each embedding dimension
  • Total dimensionality: 5 * D
  • Column naming pattern: embed_0000_mean ... embed_(D-1)_p90

Supported models and dimensions:

model_tag	HuggingFace / model identifier	Per-frame dimension
dinov2-small	facebook/dinov2-small	384
dinov2-base	facebook/dinov2-base	768
dinov2-large	facebook/dinov2-large	1024
vit-mae-base	facebook/vit-mae-base	768
vit-base-patch16-224	google/vit-base-patch16-224	768
siglip-base-patch16-224	google/siglip-base-patch16-224	768
siglip-so400m-patch14-384	google/siglip-so400m-patch14-384	1152
clip-vit-base-patch32	openai/clip-vit-base-patch32	768
clip-vit-large-patch14	openai/clip-vit-large-patch14	1024