# 10th Place Solution

First, I would like to express my gratitude for organizing this exciting competition. For me, this marks my first gold medal in nearly two years. Moreover, since I previously achieved 10th place at the same Child Mind Institute competition through shakeup, these results feel particularly meaningful to me.
Training Strategy
Epochs & Optimizer
: 200 epochs with
AdamW
optimizer and
EMA
(decay=0.999), no learning rate scheduling.
Validation Strategy
:
StratifiedGroupKFold
with
groups="subject"
,
k=5
. CV splits were generated using three different seed values: 0, 1, and 2.
Experiment Management
:
For clearly underperforming experiments, only the initial seed (seed 0) was run. All other experiments were evaluated across all three seeds.
The F1 metric showed significant instability; in some cases, performance was high in seed=0 folds but low in seeds 1 and 2.
Model Selection
:
Adopted the best epoch based on the average performance across all 15 experiments (5 folds x 3 seeds).
Additionally, if the F1 metric was unchanged but the loss improved, that model was selected.
Preprocessing
Handling of NaN values
Rotation (
rot
)
: Replaced missing values with unit quaternions.
Thermal (
thm
)
:
For each sequence, if any sensor value was below 18 degrees, we replaced that sensor's corresponding value with NaN. This was done to mask potentially corrupted readings.
Calculated the row-wise mean of NaN values and used it for imputation.
For rows containing only NaN values, we assigned a global mean of
27.03
.
Hand preference correction
Adjusted all sensor values to a right-handed orientation.
df[] *= -
df[] *= -
df[] *= -


thm_3 = df[].copy()
thm_5 = df[].copy()
df[] = thm_3
df[] = thm_5


tof_names = [  i  ()]
tof = df[tof_names].values.copy()
tof = tof.reshape(-, , , )
tof = tof[:, [, , , , ]]
tof = np.flip(tof, )
tof = tof.reshape(-,  *  * )
df[tof_names] = tof.copy()
Z-axis rotation
It was determined that
SUBJ_019262
and
SUBJ_045235
had their sensors rotated 180 degrees around the Z-axis. The following correction was applied to their data.
df[] *= -
df[] *= -


w = df[].copy()
x = df[].copy()
y = df[].copy()
z = df[].copy()
df[] = -z
df[] = y
df[] = -x
df[] = w


thm_2 = df[].copy()
thm_3 = df[].copy()
thm_4 = df[].copy()
thm_5 = df[].copy()
df[] = thm_4
df[] = thm_2
df[] = thm_5
df[] = thm_3


tof_names = [  i  ()]
tof = df[tof_names].values.copy()
tof = tof.reshape(-, , , )
tof = np.rot90(tof, k=, axes=(, ))
tof = tof[:, [, , , , ]]
tof = tof.reshape(-,  *  * )
df[tof_names] = tof.copy()
Data Augmentation
Due to the limited dataset size, data augmentation was crucial for model performance.
Time-series stretching
Quaternions(rot)
:
Problem
: Quaternions can represent the same orientation with inverted signs ($q$ and $-q$). This causes "inversion points" in sequences where interpolation-based stretching produces meaningless data.
Solution
: We preprocessed sequences to ensure all pairwise inner products were positive, creating a smooth orientation path.
Method
: Used
Slerp
or
Nearest
interpolation methods.
Acceleration(acc)
:
Problem
: Simple interpolation (e.g., stretching by 1.5x) incorrectly scales the integrated velocity and position derived from acceleration.
Solution
: We implemented a correction to preserve the start and end coordinates after stretching.
Procedure
:
Use the orientation (
rot
) data to separate each frame's acceleration into its
gravitational
and
linear
components.
Apply stretching only to the linear acceleration component, then divide it by the square of the stretching factor.
Recalculate the gravitational component using the stretched
rot
values.
Rotation
Rotating the data to simulate wristband misalignment was highly effective.
We applied a random rotational bias to the entire sequence along the Y, X, and Z axes.
The optimal standard deviations for the bias were:
std_Y=10
,
std_X=7
, and
std_Z=0.1
.
MIXUP
MIXUP
proved to be an extremely effective technique.
Implementation
:
We applied MIXUP
after
feature engineering.
To prevent overfitting and improve performance over longer training, we applied MIXUP in
multiple stages
:
z = Mixup(Mixup(Mixup(x)))
.
To prevent excessive averaging, the original data
x
was concatenated with the mixed data
Mixup(x)
along the batch dimension.
Parameters
: The first stage used
beta=0.5
, while the second and third stages used
beta=16
.
We also experimented with mixing three or more data samples by sampling from a Dirichlet distribution, but the multi-stage approach yielded better results.
Model Architecture
Input
: The model processes input sequences of
128 frames
.
Receptive Field
: Long-term temporal dependencies were not required for this task. The model's receptive field spanned
15 frames
(~1.5 seconds), which was sufficient. This was achieved with 7 convolutional layers (kernel size 3) without pooling.
Attention
: Features were weighted using
additive attention
, consolidated into a single representation, and then passed to a classifier.
Two-Stage Model
:
Stage 1
: An initial model estimates 4 posture classes and feeds these probability values to the second stage.
Stage 2
: A second model inputs both the raw sensor data and the posture probabilities to predict the final 18-class and 9-class targets.
Ensemble
: We created and used separate models: one for IMU-only data and one for all-sensor data.
Feature Engineering
IMU (12 features)
:
Acceleration, acceleration norm, linear acceleration, linear acceleration norm, angular velocity, angular velocity norm.
THM (10 features)
:
Temperature, temporal derivative of temperature.
TOF (5 features)
:
Sensor-wise averages.
Reason for Significant Performance Improvement ("Shake Up")
We achieved a substantial improvement from
24th place (Public LB) to 10th place (Private LB)
.
Primary Reason
: This improvement likely stems from the preprocessing step where we detected and corrected data from sensors that were flipped 180 degrees on the Z-axis.
Score Impact
:
Public Leaderboard:
0.874 (no change)
Private Leaderboard:
0.857 → 0.867 (0.01 improvement)
We did not perform any private probing.
Given that the sensor-flip case occurred in 2 out of 82 subjects in the training data, the probability of it being present in at least one of the private test subjects was approximately 40%. Therefore, we consider this result to be fortunate.