Publication schematics

These original SVG illustrations summarize methods, rather than reproducing paper figures or portraying invented experimental outputs. Drawn directly as vectors for precise labels and arrows; no image-generation outputs are used. Each visible schematic can be opened at full size from the publication list. Original thumbnails remain in the repository.

Sources and visual interpretation

Asset	Source read	What the schematic represents
svp.svg	arXiv:2501.04568v2, §3, Figs. 7–9; accepted TMLR record	Draft captions → grounding feedback → refined text → informative sample selection → model adaptation. Grounding is feedback during sampling, not a direct box-coordinate training target.
dehazing.svg	arXiv:2004.08554v3, §3, Fig. 2, Algorithm 1	Stereo movie frames → horizontal disparity and refined depth → physically rendered hazy/clean pairs. This is a dataset contribution, not a dehazing architecture.
reflection.svg	arXiv:1912.03623v3, §III, Figs. 2 and 4	Shared-background image pairs constrain background/reflection separation during training through cross-reconstruction. The trained network takes only one image at inference. The thumbnail compresses latent-code switching into a shared network.
laplacian.svg	arXiv:1906.08635v4, §4.4–4.5, Fig. 3	Deep features form a graph; normalized non-smooth 1-Laplacian energy infers pseudo-labels. Graph inference and classifier training alternate. Graph colors denote classes, not measured probabilities.
reflectance.svg	arXiv:2211.14751v3, Proposed Method, Fig. 2	Prior-guided initial reflectance estimation, then shadow/specular-aware attention refinement. Output is diffuse reflectance, not a relit photograph.
interpolation.svg	Publisher abstract and publicly available Fig. 2	Forward and backward warping use optical flow, contextual features and edges; the synthesis network combines their outputs into an intermediate frame.
tracking.svg	Local paper, §3.1–3.3, Figs. 1–2	Camera-compensated history drives multi-stream ConvLSTM future-location prediction. A confidence selector chooses between trajectory and appearance-based tracking when the target is occluded.
all-weather.svg	Local paper, §3, Figs. 2–4	Weather-specific encoders → architecture/feature search using physics-inspired operations → shared decoder. Categorical adversarial supervision is omitted from the compact inference schematic.
rainflow.svg	Local paper, §3.1–3.3, Fig. 2	Learned feature multipliers handle veiling; learned chromatic max/min feature mappings handle streaks. Combined matching costs support optical flow. Output arrows are illustrative flow, not a measured result.
graphxnet.svg	Local paper, §2, Figs. 1–2	A weighted graph has one node per X-ray image. A few labeled nodes constrain multi-class 1-Laplacian optimization to classify unlabeled images. The schematic does not depict pixel segmentation or a clinical diagnosis.
heavy-rain.svg	Local paper, §3.1–3.2, Figs. 2–3	Residue-guided decomposition supports estimation of streaks S, transmission T and atmospheric light A; physical reconstruction is refined by a depth-guided conditional GAN. The compact “veil” label refers to transmission-related attenuation.
robust-flow.svg	Local paper, §3–4, Eq. 12	Colored residue cues suppress achromatic rain, while smooth structure compensates for dark/achromatic regions and noise. Both inform joint variational optical flow. Max-minus-min denotes the residue channel underlying the colored representation.
multiscale-derain.svg	arXiv:1712.06830v1, §4.1–4.2, Fig. 5	Parallel recurrent subnetworks estimate small, medium and large streaks, with stage-wise refinement and a parallel veil module. DenseNet feature extraction is omitted to keep the thumbnail legible.

Access and scope

Detail Preserving Novel View Synthesis has only a placeholder # paper link. Its existing image and entry were left unchanged, as requested.
The full Hybrid Warping Fusion article was paywalled. Its diagram is limited to the publisher’s abstract and public architecture figure, which were inspected directly.
OpenReview blocked downloading the accepted SVP PDF. Its accepted title, venue, year and author order were verified from the indexed published PDF; method details were read from the linked arXiv v2 manuscript. No unverified revision-specific mechanisms are illustrated.
Source reading focused on contributions, method sections, architecture figures and relevant results. This is a visual-summary task, not a peer review or replication.
Illustrative cars, landscapes, graphs and X-ray shapes are schematic examples, not samples copied from datasets or experimental evidence.

Website content updates

The bio now lists Senior Applied Scientist at Microsoft, retaining Amazon and ByteDance as previous roles. The SVP entry uses the accepted title Feedback-Driven Vision-Language Alignment via Sampling-based Visual Projection, TMLR 2026, the OpenReview link and the published author order. Submission number 8150 is not presented as a publication identifier.