Task contract. Convert a rolling-shutter sequence with row-wise geometric distortion into the time-aligned global-shutter sequence while preserving content, framing, appearance, and motion.
4official archives
15,595,587,207S3 bytes
6S3 objects
4/4ZIP checks passed
Input → target previews
Stage 1 status. These three pairs are preliminary pipeline evidence. They are not commit-bound Gate B evidence and do not represent reviewer approval.
1. Fastec-RS test — real capture
test/190206_Tuebingen_000100 · 34 frames · 640×480 · 10 FPS · 3.4 s
INPUT: ordered
*_rolling.pngTARGET: matched
*_global_middle.png2. Fastec-RS train — real capture
train/190206_Tuebingen_000118 · 34 frames · 640×480 · 10 FPS · 3.4 s
INPUT: ordered
*_rolling.pngTARGET: matched
*_global_middle.png3. Carla-RS train — synthetic fast camera rotation
train/Carla_town03/seq_55 · 10 frames · 640×448 · 10 FPS · 1.0 s · selected for visible row-wise geometry change
INPUT: ordered
*_rs.pngTARGET: matched
*_gs_m.pngIdentity and provenance
| Dataset | Fastec-RS and Carla-RS |
|---|---|
| Authors | Peidong Liu, Zhaopeng Cui, Viktor Larsson, Marc Pollefeys |
| Paper | Deep Shutter Unrolling Network, CVPR 2020 |
| Official source | DeepUnrollNet repository, four author-published Google Drive links |
| Acquisition | Fresh download by Nancy on 2026-07-20; raw archives retained unchanged |
| S3 raw prefix | s3://tv2v-data/raw/taxonomy/Task_119_rolling_shutter_correction_data-pipeline/ |
| License | Dataset archive license not explicitly located; redistribution and commercial use remain unclear. The repository code license must not be assumed to cover data. |
Measured archive inventory
| Archive | Bytes | Contents measured with zipinfo | SHA-256 |
|---|---|---|---|
| carla_rs_test.zip | 1,993,142,592 | 4,044 files; 2,000 PNG | 4e9a9069…1117dad2 |
| carla_rs_train.zip | 9,997,166,632 | 21,328 files; 10,550 PNG | 3d08184a…fdcc4a11 |
| fastec_rs_test.zip | 974,958,671 | 1,981 files; 1,960 PNG; 21 meta logs | 93bdd04b…767a626f |
| fastec_rs_train.zip | 2,630,318,839 | 5,768 files; 5,712 PNG; 56 meta logs | 833ba4c9…7f68ce10 |
The S3 prefix also contains SHA256SUMS.txt and STATUS.json. All six S3 object sizes match the verified local sources exactly.
Raw location, acquisition, and contamination
| Official download set | Four Google Drive archives published by the DeepUnrollNet authors: Fastec train/test and Carla train/test |
|---|---|
| Internal raw location | s3://tv2v-data/raw/taxonomy/Task_119_rolling_shutter_correction_data-pipeline/ |
| Fresh download? | Yes, directly from all four official repository links on 2026-07-20 |
| Known modifications | None to ZIP contents. SHA256SUMS.txt and STATUS.json are added provenance records. |
| Registration / DUA | No registration observed; no separate data-use agreement found |
| Re-download | Official links were available at audit time |
Scale, splits, formats, and annotation semantics
| Source/item | Measured scale / format | Meaning |
|---|---|---|
| Fastec train | 56 meta-log scenes; 5,712 PNG; RGB uint8 640×480 | Rolling/global-middle/global-first triplets |
| Fastec test | 21 scenes; 1,960 PNG; RGB uint8 640×480 | Paired evaluation scenes; 653 rolling/global-middle pairs measured |
| Carla train/val | 21,328 files; 10,550 PNG; sampled masks 640×448 uint8 | Synthetic RS/GS plus flow, depth, masks, calibration, and poses |
| Carla test | 4,044 files; 2,000 PNG | Synthetic held-out scenes with auxiliary supervision |
*_rolling.png | RGB uint8 | Task input with row-wise readout distortion |
*_global_middle.png | RGB uint8, same index | Primary time-aligned target |
*_global_first.png | RGB uint8 | Alternative timing target retained in raw |
meta.log | Text capture/timing parameters | Supports exported-FPS derivation |
Native structure and conversion
scene/ ├── NNN_rolling.png → input_video.mp4 ├── NNN_global_middle.png → target_video.mp4 ├── NNN_global_first.png → retained in raw └── meta.log → timing metadata (Fastec)
The official model consumes consecutive rolling-shutter images and predicts the corresponding global-shutter image. For V2V, frames are numerically ordered within one scene and paired by the same index. Fastec metadata gives 2400 FPS source capture and a 240-frame sampling interval, yielding 10 FPS for the inspected exported sequence.
Pipeline → final four-file contract
verified ZIP → extract one split/scene → match rolling/global_middle indices → reject incomplete/short scenes → derive and record FPS → equal H.264/yuv420p input/target → prompt.txt + metadata.json → contract validation → final/taxonomy/Task_119_rolling_shutter_correction/Task_119_rolling_shutter_correction_NNNNNNNNN/
Quality gates and limitations
- Every ZIP central directory passed
unzip -tq; SHA-256 values were recorded before upload. - Input/target indices, dimensions, frame counts, FPS, and durations must match per final clip.
- Short scenes are reported and excluded from normal clip windows; scenes are never concatenated.
- Fastec derives RS from high-speed GS captures; Carla is synthetic. Neither fully represents every consumer sensor artifact.
- This report covers verified raw ingestion. Final four-file V2V samples and the production pipeline remain separate work.
Completeness and source verification
| Check | Result |
|---|---|
| Official paper/repository/four source links | Verified |
| ZIP integrity | 4/4 passed |
| SHA-256 | 4/4 recorded |
| S3 reconciliation | 6 objects; 15,595,587,207 bytes; local sizes match |
| Native semantics/formats | Verified by archive reads |
| Real paired preview | Verified, 34 frames |
| Dataset license | Unclear; release blocker |
| Production final samples | Not generated |