Reproducing the results of paper Tem-adapter (Video Question-Answering)

This repository explores the reproduction and improvement of the Tem-adapter architecture for Video Question Answering (VideoQA) using the SUTD-TrafficQA dataset. The project involves replication of results using released checkpoints, training from scratch, and extending the architecture with a custom cross-attention layer.

---

Setup

• Dataset: Download the SUTD-TrafficQA dataset and place it in the data/ folder
• Released Checkpoint: Drive Link
• Reproduced Checkpoint: Drive Link

---

Results

Replication with Released Checkpoint

Source	Validation Accuracy
Original (paper)	46.00%
Reproduced (ckpt)	46.00%

✔️ Exact match with the published results using the official checkpoint.

---

Training from Scratch

Metric	Value
Sum loss	0.127
Avg loss	0.34
CE loss	33.28
Recon loss	0.0067
Average Accuracy	98.20%
Validation Accuracy	45.37%

⚠️ Minor drop (~0.63%) from original likely due to smaller batch size and different GPU.