{ "id": 663274, "title": " 52th Place Solution: Feature Engineering & Post-Processing Improvements", "url": "/competitions/MABe-mouse-behavior-detection/discussion/663274", "author_name": "Chikuwabu", "post_date": "2025-12-17T06:05:40.708000", "votes": 14, "comment_count": 0, "views": 0, "content": "

The solution introduction image was created by Nano Banana.

\n

First, I want to thank @AmbrosM and @mattiaangeli for the excellent baseline notebook. My solution builds directly on top of it, keeping the core FPS-adaptive windowing approach and GBDT ensemble (LightGBM, XGBoost, CatBoost) while adding several feature engineering and post-processing improvements.

\n

Final Rank: 52nd

\n
\n

Baseline Overview

\n

The baseline already includes:

\n\n
\n

My Improvements

\n

1. Action-Specific Thresholds & Minimum Durations

\n

Instead of using a single threshold (0.27) for all behaviors, I defined per-action thresholds based on behavior characteristics:

\n
DEFAULT_ACTION_THRESHOLDS = {\n    'rear': 0.25,\n    'selfgroom': 0.30,\n    'attack': 0.35,\n    'chase': 0.30,\n    'approach': 0.22,\n    'sniff': 0.25,\n    # ... etc\n}\n\nDEFAULT_MIN_DURATION = {\n    'rear': 5,\n    'attack': 4,\n    'chase': 5,\n    # ... etc (in frames at 30fps, scaled by actual fps)\n}\n
\n

This helps reduce false positives for aggressive behaviors while being more sensitive to subtle social interactions.

\n
\n

2. Advanced Temporal Smoothing & Event Merging

\n

Two-stage smoothing:

\n
    \n
  1. Median filter to remove single-frame outliers
    2. \n
  2. Rolling mean for final smoothing
    3. \n
\n

Event merging: If two events of the same action are separated by a small gap (< 5 frames), they are merged into one continuous event. This handles brief detection dropouts.

\n
def merge_nearby_events(submission, max_gap_frames=5):\n    # Merge events separated by small gaps\n    ...\n
\n
\n

3. Orientation & Angle Features

\n

Single mouse:

\n\n

Mouse pairs:

\n\n
# Direction from A to B\ndir_A_to_B = np.arctan2(B_cy - A_cy, B_cx - A_cx)\n# Is A facing B?\nA_facing_B = np.cos(A_orient - dir_A_to_B)  # 1 = facing, -1 = away\n
\n
\n

4. Behavior-Specific Signature Features

\n

Rearing (rear):

\n\n

Chase (chase):

\n\n

Attack (attack):

\n\n
\n

5. Metadata Features

\n

Added lab and experimental context as features:

\n\n
LAB_ID_ENCODING = {\n    'AdaptableSnail': 0,\n    'BoisterousParrot': 1,\n    # ... etc\n}\n
\n
\n

Summary of Changes

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
CategoryBaselineMy Solution
ThresholdsFixed 0.27Per-action (0.22-0.35)
SmoothingSimple rollingMedian + rolling mean
Post-processingMin 3 framesPer-action min + event merge
Orientation featuresNoneAngular velocity, facing detection
Behavior signaturesNoneRear/chase/attack specific
MetadataNoneLab, arena, sex features
\n
\n

What Didn't Work (or wasn't used)

\n", "messages": [ { "id": 3377899, "postDate": "2025-12-17T06:05:40.707Z", "content": "

The solution introduction image was created by Nano Banana.

\n

First, I want to thank @AmbrosM and @mattiaangeli for the excellent baseline notebook. My solution builds directly on top of it, keeping the core FPS-adaptive windowing approach and GBDT ensemble (LightGBM, XGBoost, CatBoost) while adding several feature engineering and post-processing improvements.

\n

Final Rank: 52nd

\n
\n

Baseline Overview

\n

The baseline already includes:

\n\n
\n

My Improvements

\n

1. Action-Specific Thresholds & Minimum Durations

\n

Instead of using a single threshold (0.27) for all behaviors, I defined per-action thresholds based on behavior characteristics:

\n
DEFAULT_ACTION_THRESHOLDS = {\n    'rear': 0.25,\n    'selfgroom': 0.30,\n    'attack': 0.35,\n    'chase': 0.30,\n    'approach': 0.22,\n    'sniff': 0.25,\n    # ... etc\n}\n\nDEFAULT_MIN_DURATION = {\n    'rear': 5,\n    'attack': 4,\n    'chase': 5,\n    # ... etc (in frames at 30fps, scaled by actual fps)\n}\n
\n

This helps reduce false positives for aggressive behaviors while being more sensitive to subtle social interactions.

\n
\n

2. Advanced Temporal Smoothing & Event Merging

\n

Two-stage smoothing:

\n
    \n
  1. Median filter to remove single-frame outliers
    2. \n
  2. Rolling mean for final smoothing
    3. \n
\n

Event merging: If two events of the same action are separated by a small gap (< 5 frames), they are merged into one continuous event. This handles brief detection dropouts.

\n
def merge_nearby_events(submission, max_gap_frames=5):\n    # Merge events separated by small gaps\n    ...\n
\n
\n

3. Orientation & Angle Features

\n

Single mouse:

\n\n

Mouse pairs:

\n\n
# Direction from A to B\ndir_A_to_B = np.arctan2(B_cy - A_cy, B_cx - A_cx)\n# Is A facing B?\nA_facing_B = np.cos(A_orient - dir_A_to_B)  # 1 = facing, -1 = away\n
\n
\n

4. Behavior-Specific Signature Features

\n

Rearing (rear):

\n\n

Chase (chase):

\n\n

Attack (attack):

\n\n
\n

5. Metadata Features

\n

Added lab and experimental context as features:

\n\n
LAB_ID_ENCODING = {\n    'AdaptableSnail': 0,\n    'BoisterousParrot': 1,\n    # ... etc\n}\n
\n
\n

Summary of Changes

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
CategoryBaselineMy Solution
ThresholdsFixed 0.27Per-action (0.22-0.35)
SmoothingSimple rollingMedian + rolling mean
Post-processingMin 3 framesPer-action min + event merge
Orientation featuresNoneAngular velocity, facing detection
Behavior signaturesNoneRear/chase/attack specific
MetadataNoneLab, arena, sex features
\n
\n

What Didn't Work (or wasn't used)

\n", "rawMarkdown": "The solution introduction image was created by Nano Banana.\n\nFirst, I want to thank @AmbrosM and @mattiaangeli for the excellent baseline notebook. My solution builds directly on top of it, keeping the core FPS-adaptive windowing approach and GBDT ensemble (LightGBM, XGBoost, CatBoost) while adding several feature engineering and post-processing improvements.\n\n**Final Rank: 52nd**\n\n---\n\n## Baseline Overview\n\nThe baseline already includes:\n- FPS-adaptive feature scaling (`_scale` function)\n- Body part distance features\n- Rolling statistics (mean, std, range) for position and velocity\n- Curvature, multi-scale speed, and state transition features\n- GPU-accelerated GBDT ensemble\n\n---\n\n## My Improvements\n\n### 1. Action-Specific Thresholds & Minimum Durations\n\nInstead of using a single threshold (0.27) for all behaviors, I defined per-action thresholds based on behavior characteristics:\n\n```python\nDEFAULT_ACTION_THRESHOLDS = {\n 'rear': 0.25,\n 'selfgroom': 0.30,\n 'attack': 0.35,\n 'chase': 0.30,\n 'approach': 0.22,\n 'sniff': 0.25,\n # ... etc\n}\n\nDEFAULT_MIN_DURATION = {\n 'rear': 5,\n 'attack': 4,\n 'chase': 5,\n # ... etc (in frames at 30fps, scaled by actual fps)\n}\n```\n\nThis helps reduce false positives for aggressive behaviors while being more sensitive to subtle social interactions.\n\n---\n\n### 2. Advanced Temporal Smoothing & Event Merging\n\n**Two-stage smoothing:**\n1. Median filter to remove single-frame outliers\n2. Rolling mean for final smoothing\n\n**Event merging:** If two events of the same action are separated by a small gap (< 5 frames), they are merged into one continuous event. This handles brief detection dropouts.\n\n```python\ndef merge_nearby_events(submission, max_gap_frames=5):\n # Merge events separated by small gaps\n ...\n```\n\n---\n\n### 3. Orientation & Angle Features\n\n**Single mouse:**\n- Body orientation angle (nose-to-tail direction)\n- Angular velocity and acceleration\n- Rolling angular statistics\n\n**Mouse pairs:**\n- `A_facing_B` / `B_facing_A`: Is mouse A's body oriented toward B?\n- `mutual_facing`: Both mice facing each other (for sniffing)\n- `A_behind_B`: Following/chase pattern detection\n- `orient_parallel`: Parallel vs perpendicular orientation\n\n```python\n# Direction from A to B\ndir_A_to_B = np.arctan2(B_cy - A_cy, B_cx - A_cx)\n# Is A facing B?\nA_facing_B = np.cos(A_orient - dir_A_to_B) # 1 = facing, -1 = away\n```\n\n---\n\n### 4. Behavior-Specific Signature Features\n\n**Rearing (`rear`):**\n- Nose-tail distance compression (body appears shorter from top-down view)\n- `rear_ratio`: nose-to-center vs center-to-tail distance ratio\n- Body length variance (increases during rearing transitions)\n\n**Chase (`chase`):**\n- Sustained approach rate (negative distance derivative)\n- Velocity correlation between mice (high = following)\n- Speed ratio (chaser typically faster)\n\n**Attack (`attack`):**\n- Proximity indicator (very close < 3cm)\n- Rapid approach detection\n- Acceleration bursts\n\n---\n\n### 5. Metadata Features\n\nAdded lab and experimental context as features:\n\n- **Lab ID encoding**: Different labs may have different annotation styles\n- **Arena shape/size**: Square, rectangular, or circular; normalized dimensions\n- **Number of mice**\n- **FPS** (normalized)\n- **Mouse sex**: Agent/target sex, same-sex interaction flag\n\n```python\nLAB_ID_ENCODING = {\n 'AdaptableSnail': 0,\n 'BoisterousParrot': 1,\n # ... etc\n}\n```\n\n---\n\n## Summary of Changes\n\n| Category | Baseline | My Solution |\n|----------|----------|-------------|\n| Thresholds | Fixed 0.27 | Per-action (0.22-0.35) |\n| Smoothing | Simple rolling | Median + rolling mean |\n| Post-processing | Min 3 frames | Per-action min + event merge |\n| Orientation features | None | Angular velocity, facing detection |\n| Behavior signatures | None | Rear/chase/attack specific |\n| Metadata | None | Lab, arena, sex features |\n\n---\n\n## What Didn't Work (or wasn't used)\n\n- TCN sequence model: Implemented but disabled due to time constraints\n- CV-based threshold optimization: Too slow for the final submission\n\n", "votes": 14 } ], "comments": [], "raw_markdown_by_id": { "3377899": "The solution introduction image was created by Nano Banana.\n\nFirst, I want to thank @AmbrosM and @mattiaangeli for the excellent baseline notebook. My solution builds directly on top of it, keeping the core FPS-adaptive windowing approach and GBDT ensemble (LightGBM, XGBoost, CatBoost) while adding several feature engineering and post-processing improvements.\n\n**Final Rank: 52nd**\n\n---\n\n## Baseline Overview\n\nThe baseline already includes:\n- FPS-adaptive feature scaling (`_scale` function)\n- Body part distance features\n- Rolling statistics (mean, std, range) for position and velocity\n- Curvature, multi-scale speed, and state transition features\n- GPU-accelerated GBDT ensemble\n\n---\n\n## My Improvements\n\n### 1. Action-Specific Thresholds & Minimum Durations\n\nInstead of using a single threshold (0.27) for all behaviors, I defined per-action thresholds based on behavior characteristics:\n\n```python\nDEFAULT_ACTION_THRESHOLDS = {\n 'rear': 0.25,\n 'selfgroom': 0.30,\n 'attack': 0.35,\n 'chase': 0.30,\n 'approach': 0.22,\n 'sniff': 0.25,\n # ... etc\n}\n\nDEFAULT_MIN_DURATION = {\n 'rear': 5,\n 'attack': 4,\n 'chase': 5,\n # ... etc (in frames at 30fps, scaled by actual fps)\n}\n```\n\nThis helps reduce false positives for aggressive behaviors while being more sensitive to subtle social interactions.\n\n---\n\n### 2. Advanced Temporal Smoothing & Event Merging\n\n**Two-stage smoothing:**\n1. Median filter to remove single-frame outliers\n2. Rolling mean for final smoothing\n\n**Event merging:** If two events of the same action are separated by a small gap (< 5 frames), they are merged into one continuous event. This handles brief detection dropouts.\n\n```python\ndef merge_nearby_events(submission, max_gap_frames=5):\n # Merge events separated by small gaps\n ...\n```\n\n---\n\n### 3. Orientation & Angle Features\n\n**Single mouse:**\n- Body orientation angle (nose-to-tail direction)\n- Angular velocity and acceleration\n- Rolling angular statistics\n\n**Mouse pairs:**\n- `A_facing_B` / `B_facing_A`: Is mouse A's body oriented toward B?\n- `mutual_facing`: Both mice facing each other (for sniffing)\n- `A_behind_B`: Following/chase pattern detection\n- `orient_parallel`: Parallel vs perpendicular orientation\n\n```python\n# Direction from A to B\ndir_A_to_B = np.arctan2(B_cy - A_cy, B_cx - A_cx)\n# Is A facing B?\nA_facing_B = np.cos(A_orient - dir_A_to_B) # 1 = facing, -1 = away\n```\n\n---\n\n### 4. Behavior-Specific Signature Features\n\n**Rearing (`rear`):**\n- Nose-tail distance compression (body appears shorter from top-down view)\n- `rear_ratio`: nose-to-center vs center-to-tail distance ratio\n- Body length variance (increases during rearing transitions)\n\n**Chase (`chase`):**\n- Sustained approach rate (negative distance derivative)\n- Velocity correlation between mice (high = following)\n- Speed ratio (chaser typically faster)\n\n**Attack (`attack`):**\n- Proximity indicator (very close < 3cm)\n- Rapid approach detection\n- Acceleration bursts\n\n---\n\n### 5. Metadata Features\n\nAdded lab and experimental context as features:\n\n- **Lab ID encoding**: Different labs may have different annotation styles\n- **Arena shape/size**: Square, rectangular, or circular; normalized dimensions\n- **Number of mice**\n- **FPS** (normalized)\n- **Mouse sex**: Agent/target sex, same-sex interaction flag\n\n```python\nLAB_ID_ENCODING = {\n 'AdaptableSnail': 0,\n 'BoisterousParrot': 1,\n # ... etc\n}\n```\n\n---\n\n## Summary of Changes\n\n| Category | Baseline | My Solution |\n|----------|----------|-------------|\n| Thresholds | Fixed 0.27 | Per-action (0.22-0.35) |\n| Smoothing | Simple rolling | Median + rolling mean |\n| Post-processing | Min 3 frames | Per-action min + event merge |\n| Orientation features | None | Angular velocity, facing detection |\n| Behavior signatures | None | Rear/chase/attack specific |\n| Metadata | None | Lab, arena, sex features |\n\n---\n\n## What Didn't Work (or wasn't used)\n\n- TCN sequence model: Implemented but disabled due to time constraints\n- CV-based threshold optimization: Too slow for the final submission\n\n" } }