nsubedi123/EM-QnF

ReFocus: Interactive Episodic Memory with User Feedback

CVPR 2026

Nikesh Subedi¹, Loris Bazzani², Ziad Al-Halah¹

¹University of Utah    ²University of Verona

---

Overview

We introduce EM-QnF (Episodic Memory with Questions and Feedback), a new task that extends natural language query-based episodic memory retrieval to an interactive setting. Rather than relying on a single one-shot prediction, users can provide feedback on the model's initial prediction or add more information, enabling iterative refinement of results.

We also propose FALM (Feedback ALignment Module), a plug-and-play module that enables existing temporal grounding models to incorporate user feedback effectively without architectural changes.

Dataset

This repository contains the EM-QnF dataset — feedback annotations for three episodic memory benchmarks:

Dataset	Train	Val	Test
ego4d_qnf	✓	✓	—
goalstep_qnf	✓	✓	✓
hd_epic_qnf	✓	✓	✓

Data Structure

data/
├── ego4d_qnf/           # Ego4D NLQ with feedback
│   ├── train.jsonl
│   └── val.jsonl
├── goalstep_qnf/        # Ego4D-GoalStep NLQ with feedback
│   ├── train.jsonl
│   ├── val.jsonl
│   └── test.jsonl
└── hd_epic_qnf/         # HD-EPIC NLQ with feedback
    ├── train.jsonl
    ├── val.jsonl
    └── test.jsonl

Each JSONL file has one sample per line. Common fields:

Field	Description
video_id	Clip identifier
query_id	Unique annotation UID
query_type	Task type (e.g. nlq_feedback, goalstep_nlq_feedback)
query	Original natural language query
feedback	User feedback text (feedback records only)
timestamps	Ground-truth [[start, end]] in seconds (clip-relative)
clip_span	[clip_start, clip_end] within the full video
duration	Clip duration in seconds
pred_timestamp	Model's prior prediction [start, end] that the feedback refers to

Prerequisites

The EM-QnF feedback annotations build on top of the original datasets. Please obtain access to:

• Ego4D — for Ego4D NLQ
• Ego4D-GoalStep — for GoalStep NLQ
• HD-EPIC — for HD-EPIC NLQ

Code

Inference

GroundNLQ/eval_jsonl.py runs inference on any JSONL split from this dataset using a trained checkpoint. It requires LMDB video/text features (see feature extraction) and a config file.

cd GroundNLQ
# compile NMS first
cd libs/utils && python setup.py install --user && cd ../..

python eval_jsonl.py \
    --config configs/refocus_emqnf.yaml \
    --checkpoint path/to/checkpoint.pth.tar \
    --val_jsonl ../data/ego4d_qnf/val.jsonl \
    --task nlq_feedback \
    --gt_file path/to/gt.json \
    --output predictions.pkl

Supported --task values: nlq, nlq_feedback, goalstep_nlq, goalstep_nlq_feedback, hd_epic_nlq, hd_epic_nlq_feedback.

Training code will be released soon. Stay tuned!

Citation

If you use ReFocus or the EM-QnF dataset in your research, please cite:

@inproceedings{subedi2026refocus,
  title     = {Interactive Episodic Memory with User Feedback},
  author    = {Subedi, Nikesh and Bazzani, Loris and Al-Halah, Ziad},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer
               Vision and Pattern Recognition (CVPR)},
  year      = {2026},
}

License

The EM-QnF annotations are released under CC BY-NC 4.0. Use of the underlying video data is subject to the respective dataset licenses.