FMOX: Dual Roles in Vision and Excitonics

• FMOX is a context‐dependent term used in computer vision as a unified JSON benchmark that aggregates datasets to evaluate tracking of fast, blurred objects.
• In photosynthetic excitonics, FMOX serves as shorthand for studies on the Fenna–Matthews–Olson complex, covering excitation energy transfer, vibrational dynamics, and redox effects.
• The dual usage of FMOX highlights the need for clear disambiguation, with each field leveraging its methodology to enhance either tracking performance or exciton transport modeling.

FMOX is a context-dependent technical designation that appears in two unrelated research literatures. In computer vision, it denotes a unified JSON-based benchmark format and dataset aggregation for evaluating single-object tracking and segmentation pipelines on fast moving objects, with an emphasis on small, blurred, fast-displacing targets that are underrepresented in mainstream VOT and VOS benchmarks (Aktas et al., 10 Dec 2025). In photosynthetic excitonics, the same label is used in synthesized notes for the Fenna–Matthews–Olson complex and adjacent themes, including excitation energy transfer, correlated protein motion, explicit exciton–vibrational dynamics, oxidized FMO, and, in one interpretation, the extended or holo 8-site complex (Fokas et al., 2014, Gestsson et al., 2024, Yeh et al., 2014). The term is therefore not semantically stable across domains, and its meaning is fixed by disciplinary context.

1. Nomenclature and domain specificity

Across the supplied literature, “FMOX” has no single canonical definition. In the fast-moving-object literature it is explicitly defined as “FMO eXtended,” a metadata layer that harmonizes heterogeneous FMO datasets into a common schema and augments them with bounding boxes and object-size labels (Aktas et al., 8 Sep 2025). In the Fenna–Matthews–Olson literature, the label is used more loosely: one synthesis contrasts the oxidized complex with the reduced complex under the heading “FMOX” (Gestsson et al., 2024); another states that “FMOX” is plausibly interpreted as the holo-FMO variant with an extra pigment, i.e. the 8-site form (Yeh et al., 2014); another uses it as a shorthand for explicit correlated exciton–vibrational dynamics in FMO (Schulze et al., 2015).

Usage of “FMOX”	Domain	Meaning in the supplied literature
FMOX	Computer vision	Unified JSON benchmark and dataset aggregation for fast moving objects
FMOX	Photosynthetic excitonics	Shorthand for FMO-related excitation-energy-transfer studies
FMOX	Redox-dependent FMO studies	Oxidized FMO contrasted with the reduced complex
FMOX	8-site FMO interpretation	Extended or holo FMO with an extra pigment

This divergence suggests that “FMOX” functions as a local shorthand rather than a field-wide standard. In encyclopedia usage, disambiguation is therefore essential before any technical discussion.

2. FMOX as a benchmark for fast moving object tracking

In computer vision, FMOX is a unified JSON-based benchmark format and dataset aggregation for evaluating single-object tracking and segmentation pipelines on fast moving objects. It consolidates Falling Object, TbD, TbD-3D, and FMOv2, excludes the multi-object “more_balls” sequence because unique IDs are missing, and yields 46 single-object sequences in total (Aktas et al., 10 Dec 2025). The benchmark does not introduce new raw videos; instead, it harmonizes established FMO datasets under a single schema, adds object-size categorization, and defines an evaluation protocol, metrics, and split (Aktas et al., 10 Dec 2025).

The benchmark is explicitly designed to stress-test trackers under high-speed motion and motion blur. FMOv2 and TbD are identified as particularly challenging because objects are small and the overlap between consecutive bounding boxes is often near zero, which reflects large displacements and motion blur (Aktas et al., 10 Dec 2025). Ground truth is standardized as per-frame bounding boxes; frames in which the object is absent have no ground-truth box, and FMOX ground truth does not provide masks or points, even though trackers may internally predict masks (Aktas et al., 10 Dec 2025).

A second description of FMOX specifies the JSON schema in detail. Top-level entries are databases, each dataset has dataset_name, version, description, and sub_datasets, and each frame entry includes img_index, image_file_name, and annotations with bbox_xyxy, object_wh, and size_category (Aktas et al., 8 Sep 2025). Size is represented in five bins. One paper expresses these as integer classes from 0=extremely tiny to 4=large (Aktas et al., 10 Dec 2025), while another gives explicit labels and side-length thresholds for a square bounding box: extremely tiny $[1\times1,8\times8)$, tiny $[8\times8,16\times16)$, small $[16\times16,32\times32)$, medium $[32\times32,96\times96)$, and large $[96\times96,\infty)$ (Aktas et al., 8 Sep 2025).

FMOX was created because the source datasets use diverse annotation formats, including text trajectories, MATLAB files, PNG masks, and run-length encodings, which complicate reproduction and benchmarking (Aktas et al., 8 Sep 2025). The harmonized representation is therefore intended to make loaders, evaluation, and cross-dataset comparison more consistent. A direct implication is that FMOX is primarily an interoperability and benchmarking layer, not a new video corpus.

3. Evaluation protocol, trackers, and benchmark results on FMOX

The benchmarked task is single-object segmentation and tracking from a single user-provided exemplar on an initialization frame, operationalized here as the first frame in which the object appears and instantiated with the ground-truth bounding box (Aktas et al., 10 Dec 2025). Boxes were chosen over points or masks because recent works show better robustness with box templates for SAM2-family trackers (Aktas et al., 10 Dec 2025). Evaluation is frame-by-frame across the full sequence. No resets or restarts are performed, the initialization frame is omitted from scoring, frames without ground truth are excluded, and if a tracker fails to output a box on a ground-truth frame then IoU and Dice for that frame are set to zero and included in the sequence-level mean (Aktas et al., 10 Dec 2025).

The benchmark reports mean IoU and mean Dice. With prediction $M$ and ground truth $G$,

$J = \frac{|M \cap G|}{|M \cup G|}$

and

$\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$

Sequence-level scores are the means over valid frames, and the paper aggregates these across sequences using mean, median, minimum, and maximum (Aktas et al., 10 Dec 2025). Success-rate, AUC, precision, recall, and DAVIS-style boundary $F$ are defined for completeness in the paper but are not used in the benchmark (Aktas et al., 10 Dec 2025).

Four SAM2-based pipelines are benchmarked. SAM2 uses an image encoder and a 7-slot FIFO memory with the first slot fixed, outputs three mask candidates, and selects the one with highest predicted IoU; the paper identifies vulnerability to distractors and misleading stored features during occlusions (Aktas et al., 10 Dec 2025). EfficientTAM replaces the heavy SAM2 encoder with a lightweight ViT and consolidates similar memory tokens into representative tokens to reduce compute and memory (Aktas et al., 10 Dec 2025). DAM4SAM introduces a Distractor-Aware Memory split into Recent Appearance Memory and Distractor Resolving Memory, with custom update rules that avoid distractor-infected frames (Aktas et al., 10 Dec 2025). SAMURAI adds Kalman-filter motion modeling and motion-aware instance-level memory, selecting high-confidence frames using affinity and motion scores rather than a fixed window (Aktas et al., 10 Dec 2025).

The reported model settings were SAM2.1 Hiera-L for SAM2 and DAM4SAM, SAM2.1 Hiera-B+ for SAMURAI, and efficienttam_s for EfficientTAM, on an NVIDIA RTX 4090, 13th Gen Intel Core i9, 64 GB RAM, CUDA 12.4.1, Ubuntu 20.04.6 under WSL (Aktas et al., 10 Dec 2025).

Tracker	Mean mIoU	Mean mDice
SAM2	0.461	0.545
EfficientTAM	0.438	0.520
DAM4SAM	0.505	0.600
SAMURAI	0.488	0.579

DAM4SAM has the best mean and median mIoU and mDice across all 46 sequences, with median mIoU $[8\times8,16\times16)$0 and median mDice $[8\times8,16\times16)$1 (Aktas et al., 10 Dec 2025). SAMURAI is second overall and wins on FMOv2, the subset described as especially motion-challenging (Aktas et al., 10 Dec 2025). SAM2 is strongest on TbD-3D, while EfficientTAM is consistently the fastest in wall-clock time on Falling Object, FMOv2, TbD, and TbD-3D, but trails in accuracy on the hardest motion-blurred sequences (Aktas et al., 10 Dec 2025).

The per-dataset ranking is: Falling Object, DAM4SAM $[8\times8,16\times16)$2 SAM2 $[8\times8,16\times16)$3 SAMURAI $[8\times8,16\times16)$4 EfficientTAM; TbD-3D, SAM2 $[8\times8,16\times16)$5 DAM4SAM $[8\times8,16\times16)$6 SAMURAI $[8\times8,16\times16)$7 EfficientTAM; FMOv2, SAMURAI $[8\times8,16\times16)$8 DAM4SAM $[8\times8,16\times16)$9 SAM2 $[16\times16,32\times32)$0 EfficientTAM; and TbD, DAM4SAM $[16\times16,32\times32)$1 SAMURAI $[16\times16,32\times32)$2 EfficientTAM $[16\times16,32\times32)$3 SAM2 (Aktas et al., 10 Dec 2025). Typical failure modes are motion blur, temporary invisibility, tiny targets, non-overlapping consecutive boxes, distractors, and poor initialization on a highly blurred first frame (Aktas et al., 10 Dec 2025). The paper does not assess statistical significance (Aktas et al., 10 Dec 2025).

A related EfficientTAM-only study evaluates FMOX with Trajectory IoU rather than mIoU and mDice. It reports dataset-level TIoU values of $[16\times16,32\times32)$4 on Falling Objects, $[16\times16,32\times32)$5 on TbD, and $[16\times16,32\times32)$6 on TbD-3D, and emphasizes that box initialization is much more reliable than point initialization in the presence of blur (Aktas et al., 8 Sep 2025). Taken together, these studies suggest that FMOX is now used both as a general-purpose standardized benchmark and as a stress test for memory design, motion modeling, and initialization policy in SAM2-derived trackers.

4. FMOX in the Fenna–Matthews–Olson literature: structure, states, and functional setting

In the photosynthetic literature, FMO denotes the Fenna–Matthews–Olson complex, a trimeric pigment–protein complex from green sulfur bacteria that bridges energy flow between antenna systems and the reaction center (Fokas et al., 2014). Each monomer has a clam-shell architecture that sequesters seven bacteriochlorophyll $[16\times16,32\times32)$7 pigments internally and hosts an eighth near the outside or baseplate interface (Fokas et al., 2014). The complex accepts excitations primarily at pigments 1, 6, and possibly 8, and delivers them toward a low-energy sink near pigment 3 adjacent to the reaction center (Fokas et al., 2014).

The standard single-exciton Hamiltonian in site basis is

$[16\times16,32\times32)$8

or, in another formulation,

$[16\times16,32\times32)$9

with $[32\times32,96\times96)$0 or $[32\times32,96\times96)$1 the site energies and $[32\times32,96\times96)$2 the couplings (Gestsson et al., 2024, Yeh et al., 2014). A closely related excitonic Hamiltonian used for structural ensemble analysis is

$[32\times32,96\times96)$3

where $[32\times32,96\times96)$4 is the local $[32\times32,96\times96)$5 excitation on pigment $[32\times32,96\times96)$6 (Fokas et al., 2014).

One line of work interprets FMOX as the 8-site, holo-FMO form. In that usage, the extra bacteriochlorophyll 8 has the highest site energy and lies near the protein surface close to the baseplate (Yeh et al., 2014). Simulated two-dimensional electronic spectroscopy identifies two dominant downhill pathways. In the first, the apo 7-site route $[32\times32,96\times96)$7 is modified into $[32\times32,96\times96)$8 because pigment 8 injects energy into exciton 6 (Yeh et al., 2014). In the second, the route $[32\times32,96\times96)$9 is strengthened because BChl 8 increases the wavefunction overlap between excitons 4 and 5; using the overlap metric

$[96\times96,\infty)$0

the paper reports $[96\times96,\infty)$1 in apo-FMO and $[96\times96,\infty)$2 in holo-FMO (Yeh et al., 2014).

A separate transport-optimization study also treats the 8-site complex, takes BChl 8 as the physiologically relevant initial excitation site, and analyzes transfer to both BChl 3 and BChl 4 as alternative target sites in the reaction-center interface region (Oh et al., 2018). This study identifies two transport pathways: a predominantly dissipative route to BChl 3 and a partially coherent route to BChl 4, with strong coherence–efficiency correlation for the latter (Oh et al., 2018). These results suggest that the 8-site interpretation of FMOX is not merely structural; it changes pathway topology, coupling geometry, and the inferred role of the reaction-center interface.

5. Dynamical frameworks and mechanistic interpretations in FMO excitation-transfer studies

One important structural-dynamical account uses rigidity analysis with FIRST and constrained geometric sampling with FRODA to generate an athermal conformational ensemble from the $[96\times96,\infty)$3 X-ray structure of Prosthecochloris aestuarii (PDB 3EOJ) (Fokas et al., 2014). With hydrogen-bond cutoff $[96\times96,\infty)$4, 713 hydrogen-bond constraints, and 236 hydrophobic constraints, the analysis finds an unusually large rigid clam-shell cluster, generally low residue RMSF, and strong spatial correlations between excitonically coupled pigments and helices that tune site energies (Fokas et al., 2014). The standard rigidity metric $[96\times96,\infty)$5 is $[96\times96,\infty)$6 at $[96\times96,\infty)$7, which the paper uses to support the description of FMO as “exceptionally high” in rigidity (Fokas et al., 2014). The central implication is that correlated inhomogeneous disorder narrows effective uncertainty in energy gaps and couplings: $[96\times96,\infty)$8 so positive covariance among structurally linked degrees of freedom reduces the effective spread relative to independent disorder (Fokas et al., 2014).

Several open-system approaches then model how vibrational environments shape transport. A non-Markovian quantum state diffusion treatment yields the convolutionless ZOFE master equation

$[96\times96,\infty)$9

and, for the FMO monomer, reproduces hierarchical equations of motion results at $M$0 and $M$1 while allowing calculations “within about one minute on a standard PC” (Ritschel et al., 2011). This study also shows that adding a large spectral-density peak near $M$2 does not change transfer dynamics because relevant electronic transition energies lie below about $M$3, whereas redistributing coupling among structured peaks below $M$4 can change the BChl 3 population after $M$5 by up to about $M$6 at $M$7 (Ritschel et al., 2011).

Explicit correlated exciton–vibrational dynamics has been studied with ML-MCTDH in both reduced and full-network models. A 3-site model with 450 vibrational degrees of freedom identifies the $M$8–$M$9 band as responsible for sub-picosecond population relaxation and coherence dephasing, and finds that a time-dependent Hartree mean-field treatment is not applicable (Schulze et al., 2015). A subsequent 7-site and 8-site treatment with 518 and 592 modes shows coherent $G$0 exchange with about $G$1 period in the 7-site model, a dominant downhill route $G$2 in the 8-site model, and vibronic excitation around $G$3–$G$4 associated with transfer into site 3 (Schulze et al., 2016). These studies jointly indicate that structured low-frequency vibrational couplings are not a perturbative correction but a mode-selective component of the transport mechanism.

Room-temperature HEOM calculations on an 8-site monomer further predict that coherence and semi-local entanglement initially concentrate among sites $G$5, decay before about $G$6, exhibit a pronounced drop in total coherence near $G$7, and then persist mainly among pairs closer to the reaction-center pathway, notably $G$8, with additional contributions from $G$9, $J = \frac{|M \cap G|}{|M \cup G|}$0, and $J = \frac{|M \cap G|}{|M \cup G|}$1 (González-Soria et al., 2020). Another study on site-varying pigment–protein interactions shows that realistic, site-dependent reorganization energies are highly optimized: the native configuration lies in the 85th percentile for transport to BChl 3 and the 78th percentile for transport to BChl 4 under permutation tests, while the coherence–efficiency correlation for the BChl 4 route is $J = \frac{|M \cap G|}{|M \cup G|}$2 and remains $J = \frac{|M \cap G|}{|M \cup G|}$3 even when $J = \frac{|M \cap G|}{|M \cup G|}$4 (Oh et al., 2018). This supports a picture with a predominantly dissipative pathway to BChl 3 and a partially coherent pathway to BChl 4.

A different proposal identifies a new transport channel, denoted FMO8 in that work, formed by the seven intramonomeric pigments and their second-nearest intermonomeric counterpart 8 (Díaz et al., 2021). Its efficiency, measured as time-averaged population at pigment 3 over short times, is about five times higher than the standard alternative at $J = \frac{|M \cap G|}{|M \cup G|}$5, and the mechanism is attributed to an interplay between incoherent FRET from pigment 8 and coherent formation of an optimal acceptor exciton within pigments 1–7 (Díaz et al., 2021). In a separate nonequilibrium Green’s-function formulation, exciton flow in FMO is described as population-driven rather than energy-funnel-driven: $J = \frac{|M \cap G|}{|M \cup G|}$6 with reservoir populations $J = \frac{|M \cap G|}{|M \cup G|}$7 and $J = \frac{|M \cap G|}{|M \cup G|}$8 the only driving force behind transfer in that model (Daré et al., 2024). The same paper proposes a reaction-center protection mechanism based on output-current reversal once $J = \frac{|M \cap G|}{|M \cup G|}$9 (Daré et al., 2024).

6. Oxidized FMO, redox dependence, and unresolved interpretation

In one redox-dependent usage, FMOX denotes the oxidized Fenna–Matthews–Olson complex (Gestsson et al., 2024). Two-dimensional electronic spectroscopy under controlled redox conditions shows that, in the reduced complex, the direct $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$0 and indirect $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$1 pathways are comparable, whereas oxidation suppresses $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$2 and enhances $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$3 (Gestsson et al., 2024). The experimentally inferred effective rates are, in $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$4, reduced: $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$5, $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$6, $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$7; oxidized: $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$8, $\mathrm{Dice} = \frac{2|M \cap G|}{|M| + |G|}.$9, $F$0 (Gestsson et al., 2024). This redirects population flow from the direct $F$1 route toward the indirect $F$2 pathway and coincides with disappearance of long-lived quantum beats under oxidation (Gestsson et al., 2024).

Theoretical interpretation is contested. Higgins et al. had linked the observations to oxidative onsite-energy shifts detuning an excitonic gap from an underdamped bacteriochlorophyll vibrational mode near $F$3, but a non-perturbative HEOM memory-kernel study concludes that onsite energy shifts alone cannot reproduce the observed oxidative changes while also preserving absorption spectra (Gestsson et al., 2024). In the paper’s “PNAS model,” converged HEOM even predicts the wrong qualitative trend for $F$4, increasing rather than decreasing upon oxidation (Gestsson et al., 2024). An “altered environment” model with small negative shifts $F$5, $F$6, $F$7, together with changing the underdamped mode frequency from $F$8 to $F$9, restores the qualitative redirection but still fails quantitatively (Gestsson et al., 2024).

This leads to a broader methodological conclusion. Within non-perturbative treatments, oxidation appears to require combined changes in electronic parameters and local environment rather than a purely electronic detuning picture (Gestsson et al., 2024). The paper explicitly argues that commonly used FMO electronic Hamiltonians, originally fitted with second-order cumulant and Redfield theories, may need reassessment with non-perturbative schemes or first-principles methods (Gestsson et al., 2024). That conclusion also helps explain why “FMOX” remains unstable as a label in the excitonics literature: the oxidized complex, the holo 8-site complex, and the broader FMO excitation-transfer problem overlap conceptually but are not interchangeable objects.

In summary, FMOX names either a fast-moving-object benchmark in computer vision or a family of shorthand usages centered on the Fenna–Matthews–Olson complex in excitonics. In the first domain, it is a harmonized JSON benchmark that has become a stringent test bed for SAM2-based trackers, with DAM4SAM giving the best overall accuracy and EfficientTAM the best speed on the reported studies (Aktas et al., 10 Dec 2025, Aktas et al., 8 Sep 2025). In the second, it indexes a technically diverse body of work on FMO structure, vibronic dynamics, coherence, transport optimization, and redox-dependent pathway changes, with no single standardized biochemical referent (Fokas et al., 2014, Ritschel et al., 2011, Gestsson et al., 2024). The shared abbreviation therefore masks two entirely different research objects: one a machine-vision benchmark for blurred, fast, tiny targets, the other a set of theoretical and spectroscopic problems concerning excitation transport in a photosynthetic pigment–protein complex.