fdfyaytkt/ear-wacv25-tsm-rgb-resnext50_32x4d

EAR-WACV25-DAKiet-TSM

The model was presented in the paper [](https://huggingface.co/papers/2503.07821).

This model is a Temporal Shift Module (TSM) based video classification model with a resnext50_32x4d backbone.

Github Repository: https://github.com/fdfyaytkt/EAR-WACV25-DAKiet-TSM

Data

The model was trained on a combination of datasets:

• —Toyota Smarthome dataset: Used for activity recognition.
• —ETRI-Activity3D: RGB videos (specific subsets or full dataset used depending on configuration).
• —ETRI-Activity3D-LivingLab: RGB videos (specific subsets or full dataset used depending on configuration).

Two configurations are detailed below, with their respective public leaderboard scores:

Config 1 (Public Leaderboard: 0.84402)

• —Toyota Smarthome dataset
• —ETRI-Activity3D - RGB videos (RGB\_P091-P100)
• —ETRI-Activity3D-LivingLab - RGB videos (RGB(P201-P230))

Config 2 (Public Leaderboard: 0.78856)

• —Toyota Smarthome dataset
• —ETRI-Activity3D - RGB videos (full)
• —ETRI-Activity3D-LivingLab - RGB videos (full)

Running

Example training and evaluation commands are provided below. Refer to the repository for complete details and options:

Train

python main.py elderly RGB --arch resnext50_32x4d --num_segments 8 --gd 20 --lr 0.001 --wd 1e-4 --lr_steps 20 40 --epochs 100 --batch-size 4 -j 32 --dropout 0.5 --consensus_type=avg --eval-freq=1 --shift --shift_div=8 --shift_place=blockres --npb

Eval

python generate_submission.py elderly --arch=resnext50_32x4d --csv_file=submission.csv  --weights=checkpoint/TSM_elderly_RGB_resnext50_32x4d_shift8_blockres_avg_segment8_e100/ckpt.best.pth.tar --test_segments=8 --batch_size=1 --test_crops=1