MCLC-NET: Multimodal Continual Learning for Leaf Counting
Ruchi Bhatt, Pratibha Kumari, Shreya Bansal, Vedant Agnihotri, Dwarikanath Mahapatra, Mukesh Saini
cs.CV · cs.LG
Quick verdict
Proposes the first multimodal continual-learning leaf counter plus a new field RGB/depth/thermal dataset; promising results, but test-set tuning and metric inconsistencies undermine the claims.
Summary
The paper proposes MCLC (branded MCLC-NET), a rehearsal-based continual learning framework for leaf counting from RGB, depth, and thermal images, and introduces MMLC, a field-collected dataset of 6.3K+ image triplets spanning four crops captured twice daily over four weeks. Each modality is encoded by a ResNet-50, weighted by modality attention, fused via cross-attention (RGB as query), and regressed to a leaf count. Under a domain-incremental scenario, training combines experience replay with a buffer selected by a hybrid uncertainty (Monte Carlo dropout) plus diversity (minimum embedding distance) score, dual feature- and input-level replay, and feature/output distillation. Across three task orderings of MMLC (CS1 crop-wise, CS2 time-wise, CS3 mixed) and an adapted MSU-PID benchmark, MCLC reports the lowest AMSE among CL baselines (EWC, GEM, A-GEM, DER++, ER, AVQACL): 0.675/0.542/0.745 on CS1/CS2/CS3 and 0.664 on MSU-PID, averaged over three seeds, though well above the joint-training bounds (0.177-0.193 on MMLC). Main limitations: hyperparameters are selected by lowest test AMSE; BWT/FWT reporting shows internal inconsistencies (sign flips, EWC's FWT of 19.3 against AMSE 2.3, MCLC beating the cumulative upper bound on MSU-PID); no significance tests; no modality ablation; buffer weights alpha/beta unreported; no code release statement.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- MCLC is the first framework to formulate leaf counting as multimodal domain-incremental continual learning, combining replay, distillation, and uncertainty-diversity buffer selection.
- The MMLC dataset is the first real-field multimodal (RGB/depth/thermal) leaf-counting benchmark designed for CL, with 8 tasks and three task orderings.
- MCLC achieves the lowest AMSE among CL methods on all three MMLC orderings (0.675, 0.542, 0.745) and on MSU-PID (0.664), averaged over three seeds.
- Regression is the best leaf-counting paradigm for this CL setting (CVPPP test MSE 1.08), and logit-distillation methods such as DER++ are unsuitable for continual regression.
- Task ordering materially affects CL performance, and uncertainty-diversity buffer selection outperforms random replay at equal memory.
Strengths
New real-field multimodal benchmark with CL-ready task design
MMLC (Sec. III) provides 6.3K+ RGB/depth/thermal triplets from four crops (capsicum, zucchini, cucumber, cauliflower) captured twice daily for four weeks with consumer/prosumer sensors (Pixel 5, RealSense D435, Fluke TiX580), partitioned into 8 crop-by-time tasks and three orderings (CS1/CS2/CS3) — addressing the controlled-environment limitation of MSU-PID.
First and reasonably thorough evaluation of CL for leaf counting
Sec. V covers six CL baselines (EWC, GEM, A-GEM, DER++, ER, AVQACL) plus naive/cumulative/joint bounds, across three orderings on MMLC (Tables V, IX) and a second dataset (Tables VI, X), with three seeds and per-buffer-size sweeps.
Honest reporting of negative results and the gap to upper bounds
Sensible adaptation of buffer selection to regression
Practicality documented
Weaknesses
Hyperparameters tuned on the test metric
Internal inconsistencies in CL metrics
No statistical support for headline margins
Central multimodal premise is never ablated
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MCLC-NET and what problem does it solve?
MCLC-NET (called MCLC in the paper) is a deep learning framework that counts leaves on plants from combined RGB, depth, and thermal images while learning continuously from a sequence of data batches (different crops, times of day, conditions) instead of one fixed dataset. It uses a small memory buffer of past samples, uncertainty- and diversity-based sample selection, and knowledge distillation to avoid catastrophic forgetting of earlier crops or conditions.
What is the MMLC dataset?
MMLC is a field-collected multimodal leaf-counting dataset introduced by this paper: over 6,300 paired RGB, depth, and thermal images of four vegetable crops (capsicum, zucchini, cucumber, cauliflower) photographed from above twice daily (morning and evening) during the first four weeks of growth. It is organized into 8 tasks (crop x capture-time) with three benchmark orderings, and the paper states it is publicly available.
How much better is MCLC than existing continual learning methods?
On the authors' benchmark, MCLC achieves the lowest average mean squared error among compared continual learning methods: 0.675 on the crop-wise ordering, 0.542 on the time-wise ordering, and 0.745 on the mixed ordering of MMLC, plus 0.664 on the adapted MSU-PID dataset, versus e.g. 0.843/0.743/0.749 for plain experience replay. Note the margins over the strongest baseline are small in the mixed ordering, results are averaged over only three runs, and no statistical significance tests are reported.
Is code or data available for this paper?
The paper states that the MMLC dataset is publicly available via a link in the abstract, but it does not mention releasing source code, and some implementation details (e.g., the uncertainty/diversity weighting coefficients) are not reported, which limits reproducibility of the full method.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are not analyzable from the extracted text (captions and axis labels only), so qualitative claims and training-curve behavior in Sec. V-E/F could not be verified.
- The MMLC dataset link renders as an unresolved placeholder ('MMLC Dataset') in the extracted text; the actual repository and its contents could not be checked.
- This is an arXiv v1 preprint (16 Sep 2026) and has apparently not yet completed peer review.
- A few table values show apparent typos in the source (e.g., '0.8610' in Table VIII, '0.157±0.32' for cumulative CS3); they are reported as printed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.