Peekaboo-VL: Camera-Motion Text Hints for Drone Video Question Answering

Peekaboo-VL is a drone (UAV) video question-answering model built on Qwen3.5-9B. It reaches 83.55% accuracy on the ReVA test leaderboard, 3.5 points above the organizers' ReMoSense model (80.1%), using one LoRA adapter, a one-sentence camera-motion hint, and letter scoring. It was built for the ReVA Drone Video Understanding Challenge at the MAVU Workshop, WACV 2027.

Code	github.com/LakoreAI/peekaboo-vl
Weights	huggingface.co/LakoreAI/peekaboo-vl (LoRA adapter, PEFT format)
Report	peekaboo_report.pdf
Benchmark	ReVA (paper, data)
Base model	Qwen/Qwen3.5-9B

Results

Split	Questions	Accuracy
ReVA test leaderboard	4,000	83.55%
Development split (seen videos, unseen questions)	2,078	85.37%
ReMoSense (organizers' baseline), test	4,000	80.1%

The hardest task is Temporal Grounding (about 70% on the development split). Most of its errors miss the right answer by about one second, the gap between two sampled frames.

Why drone video is hard, and what Peekaboo-VL does about it

Drone footage is hard for video language models: the camera never stops moving, objects are a few pixels wide, and many questions ask when something happens. ReMoSense handles camera motion by learning new motion tokens. Peekaboo-VL takes a simpler route: it measures the motion and tells the model in words.

• Camera-motion hint. A CPU probe (ORB features + RANSAC) estimates the frame-to-frame camera motion in milliseconds and writes it as one sentence in the benchmark's own answer vocabulary, for example "Camera motion: 0.0-4.1s moving forward, slow; 4.1-15.0s panning left". No new network parts are added.
• True timestamps. 32 frames per clip at 640×352, each tagged with its real time in seconds. During training, frames are sampled at a random position inside each time segment.
• Time bar. On Temporal Grounding questions, a progress bar is drawn on each frame so the model can see where in the clip it is.
• Letter scoring, not text generation. The answer is read from the model's next-token scores for A, B, C and D in a single forward pass, and training is cross-entropy over those four scores.
• PriDe debiasing. At test time, PriDe estimates the model's preference for each letter on 5% of the questions with rotated option orders, then removes it.

drone clip (15 s, 640x360)
   ├─> frame cache: 32 frames + true timestamps
   └─> camera-motion probe (ORB + RANSAC) -> "Camera motion: 0.0-4.1s moving forward ..."
          │
          v
   frames + motion sentence + question + options A-D
          -> Qwen3.5-9B + LoRA -> scores for A / B / C / D -> PriDe -> answer

Use the model

With PEFT

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel

base = AutoModelForImageTextToText.from_pretrained(
    "Qwen/Qwen3.5-9B", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "LakoreAI/peekaboo-vl")
processor = AutoProcessor.from_pretrained("Qwen/Qwen3.5-9B")

The model was trained to pick a letter, not to write free text. To get the reported scores, build the prompt the same way (32 frames with timestamps, the camera-motion sentence, the question, options A–D) and compare the next-token scores of A, B, C and D. The project code below does this for you.

With the project code

Requires Python 3.12+ and uv.

git clone https://github.com/LakoreAI/peekaboo-vl && cd peekaboo-vl
uv sync --extra llm --extra rich

# data: download ReVA, build the frame cache and the camera-motion hints
uv run python scripts/data/fetch_data.py
uv run python scripts/data/make_splits.py
uv run python scripts/data/extract_frames.py --workers 16
uv run python scripts/data/compute_motion_cues.py --workers 16

# score the test set with the released weights, then package for Codabench
hf download LakoreAI/peekaboo-vl --local-dir weights/peekaboo-vl
uv run python -m src.pipelines.predict --ckpt weights/peekaboo-vl \
    --config configs/train.yaml --split data/raw/reva/test.json \
    --tta pride --branch 8 --out_dir results/peekaboo-vl/test
uv run python scripts/submission/make_submission.py \
    --pred results/peekaboo-vl/test/predictions.json --out results/submissions/peekaboo-vl

Train it yourself

uv run python scripts/training/train.py --config configs/train.yaml

This trains the released recipe: LoRA rank 64 (alpha 128) on all language-model linear layers plus the vision-language merger, with the vision tower frozen. It runs 1 epoch over 14,068 ReVA training questions at learning rate 1e-4 (warm-up and cosine decay), in bf16. That took about 17 hours on one NVIDIA A40 (48 GB). Run uv run pytest for the offline test suite, and uv run python scripts/training/smoke_test.py for a quick end-to-end check that needs no GPU and no data.

Repository layout

src/            model, prompting, frame cache, cues, decoding, training and inference pipelines
configs/        train.yaml: the released training recipe
scripts/        data download and preparation, training, evaluation, Codabench packaging
tests/          offline pytest suite
notebooks/      self-contained Colab notebooks
docs/reports/   the technical report (LaTeX sources + PDF)

FAQ

What is ReVA? A benchmark of multiple-choice questions about short drone videos (VisDrone, UAVDT, ERA and others), covering 11 tasks such as Temporal Grounding, Change Detection, Object and Land Cover Recognition, and Trend and Pattern analysis.

Does it work on drone videos outside ReVA? It has not been tested there. The ReVA test set asks new questions about videos that also appear in training, so the scores show performance in that setting, not on unseen footage.

Why a text hint instead of a motion module? It needs no new parameters, runs on a CPU, and can be read and checked by a person. In our tests, how the hint is worded mattered more than how precise the measurement behind it was: writing it in the benchmark's answer words helped, and writing it in image terms hurt.

How low can a model score without the video? A text-only model that never sees a frame scores 55.9% on ReVA, so the practical floor is much higher than the 25% chance level.

Limitations

• Results are for the ReVA setting, where test questions are about videos seen in training.
• Temporal Grounding remains the weakest task.
• PriDe uses 5% of the test questions to estimate the letter preference. Scoring each question on its own works too, but is slightly less accurate.

Citation

@misc{le2026peekaboovl,
  title        = {Peekaboo-VL: Camera-Motion Text Hints and LoRA Letter Scoring
                  for Drone Video Question Answering},
  author       = {Le, Duc Minh},
  year         = {2026},
  howpublished = {\url{https://github.com/LakoreAI/peekaboo-vl}},
  note         = {ReVA Drone Video Understanding Challenge, MAVU Workshop @ WACV 2027}
}

License

The code is released under the MIT License. The adapter weights follow the license of the base model, Qwen/Qwen3.5-9B. The ReVA videos and annotations are for non-commercial research only and may not be redistributed; please cite the ReVA paper if you use them.