These lessons synthesize selected inspected examples. They are editorial guidance, not frequency estimates, official requirements or causal explanations of acceptance. The type guides supply drafting plans, while the atlas reports measured denominators.
Define the available observations and required outputs before explaining the model. Hidden supervision, location metadata or pretrained modules can change the scientific comparison.
Add a short training-versus-inference inventory, then ensure the architecture uses the same names and explicitly marks frozen or learned components.
Makes the input contract central: jointly estimate month/hour and location from an image, without supplying GPS at inference.
Abstract and Sections 1–2, physical pp.1–3 · Inspected paper
Asks whether late fusion is intrinsically preferable when all modalities are trained from scratch, rather than treating pretrained backbones as a neutral baseline.
Abstract and Section 1, physical pp.1–2 · Inspected paper
Boundary: A clear diagram does not show that two methods have matched supervision or training budgets; the protocol must establish that separately.
A small worked construction can make an unfamiliar representation or state update intelligible before the full system appears.
Choose a minimal example with two objects or a few time steps. Label the operation that changes, then reconnect it to the full method and its assumptions.
The overview uses three modality colors and frozen/trainable icons; the inset explains why a wrapped temporal distance differs from Euclidean distance.
Figures 2–3, p.4 · Inspected paper
A two-primitive cartoon is followed by sorted entry/exit events and piecewise density intervals; an epipolar-image diagnostic makes discontinuities visible under camera movement.
Figures 2–3, p.3 · Inspected paper
Boundary: An explanatory schematic is not a measured result. State omitted physics or computation rather than implying the cartoon is the whole algorithm.
A good comparison table reveals which conditions are comparable and where a method is not best. Aggregate performance and per-task superiority are different claims.
Group metrics by meaning and units, state model/data budgets, and highlight individual cells rather than treating an entire row as a universal winner.
The downstream table states 1.5B scale and 300B pretraining tokens, separates accuracy from CIDEr, and retains tasks where the overall-leading method is not best.
Table 4 and Section 5, p.7 · Inspected paper
Separates reimplemented prior methods and CLIP-backbone variants; data-size and label-noise tables expose distinct robustness interventions.
Table 1, p.6; Tables 4–5, p.7 · Inspected paper
Boundary: Selected rows or tasks are illustrations, not a reconstructed overall ranking. Do not invent uncertainty when a source provides none.
A resource claim requires the workload, measurement boundary and quality operating point. Disk size, peak GPU memory, training time and inference latency are not interchangeable.
Place quality beside runtime and memory, identify hardware and timing scope, and preserve out-of-memory or slower operating points where relevant.
The main table marks extra training datasets, preserves motion-region breakdowns, and places runtime and memory beside endpoint error for identical full-resolution inputs.
Table 5, p.7; evaluation protocol, pp.5–6 · Inspected paper
The caption defines Mem. as finished-model disk size, preventing readers from assuming GPU peak memory. The text distinguishes test-resolution FPS from the 720p headline.
Table 1, p.6; Section 6.3, p.7 · Inspected paper
Boundary: Reported timings on different hardware are not a matched speed comparison. Missing runtime hardware remains unknown rather than being inferred from training hardware.
The main paper should define the method and evaluation; supporting files should expose the details needed to verify them.
Use specific appendix or supplement-section references for calibration, implementation, full tables and failure videos. Keep the principal assumptions and comparison contract in the main paper.
Implementation, alternative vision-encoder allocations, scaling derivations and extrapolation discussion are routed to Appendices A–D; only main-paper pointers were inspected.
Sections 2.3, 3.1, 4.2 and 7, pp.3–8 · Inspected paper
Additional resizing-strategy results and complexity breakdowns are delegated to supplemental material; those references, not separate supplemental content, were inspected.
Section 4, pp.6–7 · Inspected paper
Boundary: These studies inspected main-text pointers, not the separate supplementary contents. A pointer does not establish that the referenced evidence is sufficient.
A polished visual can still carry an incorrect direction arrow or an overbroad caption. Final review should compare prose, axes, units and individual numerical cells.
Ask a reader to restate each plotted claim without the surrounding prose. Check that direction arrows, baseline names and restrictions agree across text, table and caption.
The definition of TPS gives its range and perfect/worst endpoints, but Table 3 reverses the hour-error and TPS direction arrows relative to the definitions and neighboring tables.
Equation 11, p.6; Table 3, p.7 · Inspected paper
Missing points are explicitly attributed to out-of-memory errors, and the limitations name textureless high-resolution holes in the MobileNet variant.
Figure 4, p.6; Section 4.3, p.8 · Inspected paper
Boundary: A located writing inconsistency is not by itself a finding that the underlying scientific method is invalid.
