BVI-VFI: Video Frame Interpolation Benchmark
- BVI-VFI is a specialized video quality dataset created to evaluate interpolation-induced distortions using subjective DMOS ratings and controlled full-reference protocols.
- It comprises 540 distorted video sequences generated from 36 diverse source videos at multiple resolutions and frame rates, covering a range of motion and texture challenges.
- The dataset highlights the limitations of traditional quality metrics on interpolated content and drives the development of tailored VFI evaluation methods.
Searching arXiv for papers on BVI-VFI and related VFI quality datasets. BVI‑VFI is a video quality database for video frame interpolation (VFI) designed to support subjective and objective assessment of interpolated video quality. It was introduced to address the gap between rapid progress in VFI algorithms and the limited understanding of how humans perceive interpolation artefacts, as well as the limited suitability of generic quality metrics for this domain. In its expanded form, the database contains 540 distorted sequences generated by applying five commonly used VFI algorithms to 36 diverse source videos with various spatial resolutions and frame rates, together with Differential Mean Opinion Scores (DMOS) derived from a large-scale subjective study involving 189 human subjects (Danier et al., 2022). Earlier reporting described a smaller configuration with 36 reference sequences and 180 distorted videos at HD resolution, reflecting the initial subjective quality study that established the need for a dedicated VFI quality benchmark (Danier et al., 2022).
1. Definition, scope, and historical development
BVI‑VFI was developed specifically for perceptual quality assessment of video frame interpolation outputs. Its core design principle is to isolate distortions induced by interpolation itself rather than by compression or unrelated degradations. The database therefore provides controlled high-frame-rate reference content and systematically generated interpolated counterparts, enabling direct study of artefacts such as motion judder, blur, ghosting, warping, temporal flicker, and failures around occlusions and motion boundaries (Danier et al., 2022).
The dataset appears in two closely related stages in the literature. The earlier subjective study describes BVI‑VFI as containing 36 reference sequences at three frame rates and 180 distorted videos generated using five VFI algorithms, with subjective opinion scores collected from 60 participants under standardized lab conditions (Danier et al., 2022). The later and more comprehensive database paper expands this into a larger benchmark with 36 source videos, 108 references derived across frame-rate conditions, and 540 distorted sequences, accompanied by more than 10,800 ratings from 189 paid subjects (Danier et al., 2022). This progression suggests that the 180-sequence study functioned as an initial benchmark, while the later release generalized the database across multiple spatial resolutions and a substantially larger subjective experiment.
A later metric-evaluation paper characterizes BVI‑VFI as a subjective video quality dataset curated for full-reference evaluation of VFI, and specifically notes that it provides uncompressed video sequences with distortions induced exclusively by VFI algorithms and DMOS labels (Daly et al., 1 Oct 2025). That paper also states that BVI‑VFI is publicly available and treats it as the most suitable benchmark in its study for full-reference VFI quality evaluation.
2. Dataset composition and generation protocol
In the expanded database, BVI‑VFI contains 36 source videos captured at 120 fps and spanning three spatial resolutions: 12 sequences at , 12 at , and 12 at . Each source clip is 5 seconds long, stored in YUV 4:2:0, 8 bit format (Danier et al., 2022). The database covers three frame rates—30, 60, and 120 fps—and was designed to include diverse motion patterns, dynamic textures, occlusions, and varied spatio-temporal activity.
The reference sequences are produced by temporal downsampling of the 120 fps source videos to 60 fps and 30 fps using frame dropping rather than frame averaging. This design choice is explicitly motivated by the desire to avoid ghosting and extra blur associated with frame averaging, thereby creating a cleaner basis for evaluating interpolation quality (Danier et al., 2022). For each reference sequence, the frame rate is halved by dropping every second frame, and the missing frames are then reconstructed through 2 interpolation back to the original reference frame rate.
The five VFI algorithms used to generate distorted content span simple baselines and deep models:
| Component | Reported variants | Characterization |
|---|---|---|
| Non-DL baselines | Frame repeating; frame averaging | Judder for repeating; blur for averaging |
| Deep methods | DVF; QVI; ST‑MFNet | Flow- and kernel-based DL methods |
Frame repeating displays the previous frame and introduces motion judder, whereas frame averaging synthesizes the middle frame by averaging two consecutive frames and introduces motion blur (Danier et al., 2022). DVF is described as a flow-based deep model assuming symmetric linear motion; QVI models second-order motion; ST‑MFNet is a kernel-based deep model predicting interpolation kernels for frame synthesis (Danier et al., 2022). In the earlier study, the same five-method structure appears, but the dataset is restricted to HD sequences derived from 12 source sequences, yielding 36 references and 180 distorted videos (Danier et al., 2022).
The later paper also reports that UHD-1 content was cropped to HD for subjective display using 9 candidate crops per UHD video, with the selected crop chosen to best preserve motion and texture characteristics measured via motion vector magnitude and dynamic texture parameter features (Danier et al., 2022).
3. Content selection and diversity characteristics
The content selection strategy was feature-guided. In the expanded dataset, five descriptors were computed per candidate source: Motion Vector magnitude (MV), Dynamic Texture Parameter (DTP), Spatial Information (SI), Temporal Information (TI), and Colourfulness (CF). The reported uniformity and range statistics over the 36 selected sources are: uniformity SI 0.93, TI 0.85, CF 0.95, DTP 0.87, MV 0.87; range SI 0.88, TI 0.97, CF 0.74, DTP 0.99, MV 0.99 (Danier et al., 2022). These values were used to ensure broad coverage of content factors relevant to interpolation difficulty and perceptual sensitivity.
The database is explicitly intended to cover challenging VFI scenarios, including large motions, dynamic textures such as foliage, water, fire, and smoke, occlusions, and varied spatio-temporal activity (Danier et al., 2022). The earlier study similarly reports diversity in motion types and dynamic textures, and states that source sequences were selected to ensure high coverage and uniformity in SI, TI, MV, and DTP (Danier et al., 2022). Example content names reported there include Bobblehead, Books, Bouncyball, Catch_track, Cyclist, Golf_side, Hamster, Lamppost, Plasma, Pond, Sparkler, and Water_splashing (Danier et al., 2022).
The later analysis further links content characteristics to perceptual outcomes. It reports that jointly higher DTP and MV/TI values led to degraded VFI quality, while single features alone were less reliable indicators of difficulty (Danier et al., 2022). This implies that BVI‑VFI is not merely diverse in a descriptive sense; it was assembled to support systematic analysis of how motion complexity and dynamic texture interact with interpolation artefacts.
4. Subjective experiment and annotation methodology
BVI‑VFI’s principal annotations are subjective quality scores expressed as DMOS. In the expanded study, the subjective experiment followed DSCQS under ITU‑R BT.500-14 compliant laboratory conditions. The display was a BENQ XL2720Z running at and 120 fps, with screen size mm and viewing distance 1008 mm. HD and cropped UHD-1 videos were viewed at picture height, while 540p videos were viewed at picture height (Danier et al., 2022). The earlier study reports closely related lab conditions under ITU‑R BT.500, also using the BENQ XL2720Z at , with a viewing distance of 1008 mm (Danier et al., 2022).
Under DSCQS, each trial presents one distorted sequence and its reference in randomized A/B order; subjects rate both sequences using continuous sliders marked at 0, 25, 50, 75, and 100, corresponding to Bad, Poor, Fair, Good, and Excellent (Danier et al., 2022). In the expanded study, the sequences were shown twice in the order A, B, A, B with 2-second grey screens between presentations, and there was no time limit on scoring. Subjects were screened using Snellen and Ishihara tests, and sessions lasted about 30 minutes (Danier et al., 2022).
The rating-to-DMOS pipeline is defined as follows. For subject 0 and distorted video 1, the differential score is
2
The DMOS for video 3 is then
4
Lower DMOS indicates higher perceived quality (Danier et al., 2022). The expanded study further applies the ITU‑T P.910 Alternating Projection method to estimate subject bias and inconsistency, remove bias, and re-weight subject contributions by consistency before producing final DMOS (Danier et al., 2022). By contrast, the earlier study reports differential scoring and mean aggregation into DMOS but does not describe AP-based debiasing (Danier et al., 2022).
The scale of the expanded subjective experiment is substantial: 189 subjects, more than 10,800 ratings, and at least 20 ratings per distorted video (Danier et al., 2022). A later evaluation paper reports the same 189-participant total and states that DMOS is available in both conventional and ITU‑T P.910 de-biased forms, using the P.910 de-biased version for metric evaluation (Daly et al., 1 Oct 2025).
Consistency analyses reported for the expanded dataset are unusually detailed. Inter-subject consistency, measured as SRCC between DMOS derived from two randomly split equal-size subject groups across 1000 splits, achieved median SRCC 0.910 with standard deviation 0.006. Intra-subject consistency, measured as SRCC between each subject’s differential scores and group DMOS, achieved median SRCC 0.712 with standard deviation 0.119 (Danier et al., 2022). These figures indicate that the subjective labels are statistically reliable within the study design.
5. Benchmark role in objective quality assessment
A central purpose of BVI‑VFI is benchmarking objective quality metrics for interpolated video. The expanded database paper evaluates 33 full-reference, reduced-reference, and no-reference image/video quality metrics on the dataset and concludes that none achieves strong correlation with perceived quality on the whole dataset (Danier et al., 2022). The best overall SRCC reported there is 0.70 for FAST, with PLCC 0.63 and RMSE 14.54 (Danier et al., 2022). Popular image metrics perform only moderately: PSNR achieves SRCC 0.65 and PLCC 0.58; FloLPIPS achieves SRCC 0.61 and PLCC 0.58 (Danier et al., 2022). No-reference models perform substantially worse overall.
The earlier 180-sequence study reaches a similar conclusion with a smaller metric set. On the full dataset, LPIPS is the best performer among the eight compared metrics, but only reaches SROCC 0.599 and PLCC 0.597, while PSNR reaches SROCC 0.520 and PLCC 0.471, and SSIM reaches SROCC 0.581 and PLCC 0.475 (Danier et al., 2022). The study therefore argues that existing metrics provide poor correlation with perceived quality on interpolated content.
The benchmarking protocol in the expanded paper uses SRCC, KRCC, PLCC, and RMSE, with a standard 5-parameter logistic mapping fit before PLCC and RMSE:
5
It also employs F-tests at 95% confidence on residuals to assess statistical superiority or equivalence among metrics (Danier et al., 2022). The earlier study likewise recommends logistic mapping before PLCC, OR, and RMSE, and pairwise F-tests on residuals (Danier et al., 2022).
A later paper introducing 6 uses BVI‑VFI as its principal full-reference evaluation set and reports that 7 improves over FloLPIPS on this dataset. In that study, FloLPIPS achieves overall PLCC 0.58, KRCC 0.40, SRCC 0.58, RMSE 14.98, whereas 8 achieves PLCC 0.67, KRCC 0.45, SRCC 0.63, RMSE 13.60, corresponding to improvements of 9 PLCC, 0 SRCC, 1 KRCC, and 2 RMSE (Daly et al., 1 Oct 2025). That same paper reports approximately 3 faster runtime and 4 lower memory usage than FloLPIPS on 1080p frames (Daly et al., 1 Oct 2025). This later use confirms BVI‑VFI’s role as a reference benchmark for developing VFI-specific perceptual metrics.
6. Findings about VFI algorithms, artefacts, and perceptual factors
BVI‑VFI was not only assembled to rank metrics; it was also used to analyze how VFI algorithms and viewing conditions affect perceived quality. The database supports comparisons across algorithm families, frame rates, and resolutions. In the expanded study, ST‑MFNet and QVI generally achieve lower DMOS than DVF and the non-DL baselines, especially at lower frame rates (Danier et al., 2022). DVF performs worst among the deep-learning methods, and this is attributed to its symmetric linear motion assumption failing under the non-linear motions present in the database (Danier et al., 2022). The earlier study similarly reports that ST‑MFNet shows the best overall performance across frame rates, that QVI performs similarly well at 120 fps, and that DVF yields the highest DMOS among the deep methods (Danier et al., 2022).
The dataset also reveals strong dependence on artefact type. Frame repeating primarily produces motion judder, and frame averaging produces blur (Danier et al., 2022, Danier et al., 2022). Deep flow-based methods produce warping errors, ghosting or haloing near edges, and failures under non-linear motion or occlusions; kernel-based methods may exhibit blurring, ringing from kernel mismatch, residual warping, and temporal inconsistencies in dynamic textures (Danier et al., 2022). The earlier study notes that artefacts in deep-learning methods can be concentrated in small areas with fast-moving foreground objects against static backgrounds, which causes failures for metrics based on arithmetic mean spatial pooling (Danier et al., 2022).
Frame rate is another important factor. The expanded paper reports that many metrics correlate best at 60 fps and worst at 30 fps, indicating limited robustness to frame-rate-dependent perceptual changes (Danier et al., 2022). It also reports that the perceived quality advantage of deep VFI over simple repeating or averaging decreases as frame rate increases; at sufficiently high frame rates, the benefit of complex interpolation may become negligible relative to simple baselines, although the paper states that this requires further study (Danier et al., 2022).
Spatial resolution affects perceptual quality as well. For most algorithms except DVF, perceived quality tended to decrease at higher spatial resolutions, though the paper notes that content differences across resolution sets also influenced this trend (Danier et al., 2022). This suggests that BVI‑VFI can be used to study interactions between spatial detail, motion complexity, and interpolation artefacts rather than only aggregate quality scores.
7. Availability, usage, and interpretive caveats
BVI‑VFI is publicly available. The earlier study provides access via https://danier97.github.io/BVI-VFI/ (Danier et al., 2022), while the expanded paper provides access via https://github.com/danier97/BVI‑VFI‑database and states that the repository includes the sequences and documentation (Danier et al., 2022). Both sources note that the dataset is intended to facilitate research on VFI quality assessment.
For benchmarking, the reported practical protocol is to use the provided references and distorted sequences, compute objective scores per distorted sequence against its corresponding reference, apply logistic mapping before PLCC and RMSE, and report SRCC, KRCC, PLCC, and RMSE against DMOS (Danier et al., 2022). For learning-based metrics, the expanded study additionally recommends content-wise train/test splits, such as 80/20 non-overlapping source splits, to test generalization to unseen content (Danier et al., 2022). The later 5 paper follows VQEG recommendations by fitting a 4-parameter logistic function before computing PLCC, SRCC, KRCC, and RMSE on BVI‑VFI (Daly et al., 1 Oct 2025).
Several caveats are explicit in the literature. First, the database exists in both an earlier 180-distortion HD form and a later 540-distortion multi-resolution form; both are called BVI‑VFI, so quantitative claims must be tied to the corresponding publication (Danier et al., 2022, Danier et al., 2022). Second, a later metric paper reports “180 sequences across multiple frame rates, resolutions and interpolation methods” in its abstract, but its detailed experimental description also states that its evaluated subset contains 240 scored sequences from 24 sources, 5 methods, and 2 frame-rate targets, and it does not reconcile that discrepancy (Daly et al., 1 Oct 2025). Third, BVI‑VFI should not be conflated with other VFI-related datasets. In particular, the VFI difficulty assessment dataset introduced for dynamic routing in interpolation pipelines is a separate resource for difficulty pre-assessment rather than subjective video quality benchmarking (Chen et al., 2023), and the single-frame perceptual dataset built from AIM 2020 VTSR clips is explicitly stated not to be BVI‑VFI (Han et al., 2023).
Taken together, the literature positions BVI‑VFI as a specialized benchmark for subjective and objective evaluation of interpolation quality, with strong emphasis on pure VFI-induced distortions, controlled frame-rate manipulation, multi-resolution coverage, and reliable DMOS labels. Its principal scientific significance lies in demonstrating that commonly used quality metrics, including PSNR, SSIM, and LPIPS, are not sufficient proxies for human perception on interpolated content, thereby motivating the development of bespoke VFI quality metrics (Danier et al., 2022, Danier et al., 2022).