AnimalHaze3k: Synthetic Hazy Wildlife Dataset

• AnimalHaze3k is a paired dataset comprising 1,159 clear images and 3,477 synthetic hazy variants that enables supervised dehazing research in realistic wildlife contexts.
• The dataset is generated using a physics-based haze synthesis pipeline with depth estimation and variable haze parameters, ensuring diverse and controlled haze conditions.
• Coupled with the IncepDehazeGan model, AnimalHaze3k serves both as a dehazing benchmark (achieving high SSIM and PSNR scores) and as a tool to assess improvements in downstream animal detection.

Searching arXiv for the cited paper and directly related sources. AnimalHaze3k is a synthetic hazy-wildlife image dataset designed for single-image dehazing and wildlife computer vision tasks, introduced together with the IncepDehazeGan architecture as a joint contribution for wildlife imagery acquired under atmospheric haze (Rai et al., 17 Apr 2026). It comprises 1,159 clear wildlife photographs and 3,477 hazy images generated from those clear photographs through a physics-based pipeline, with every hazy image paired to a pixel-aligned clear counterpart. The benchmark is purpose-built for supervised dehazing and for measuring the effect of haze and dehazing on downstream animal detection in conservation settings.

1. Dataset definition and composition

AnimalHaze3k is defined as a paired synthetic-hazy / real-clear wildlife dataset. The dataset contains 1,159 clear wildlife photos and 3,477 hazy images, with each clear image producing 3 distinct hazy variants (Rai et al., 17 Apr 2026). The train/validation/test partition is also explicit: the training split contains 1,041 clear and 3,123 hazy images, the validation split contains 59 clear and 177 hazy images, and the test split contains 59 clear and 177 hazy images.

Split	Clear images	Hazy images
Training	1,041	3,123
Validation	59	177
Test	59	177

The central annotation in AnimalHaze3k is the paired structure itself: each hazy image has a pixel-aligned clear image as ground truth. This is the key property that enables supervised dehazing. The paper does not mention bounding boxes, species labels attached directly to AnimalHaze3k, segmentation masks, keypoints, or behavior labels. In that sense, AnimalHaze3k is primarily a dehazing benchmark rather than a general-purpose wildlife annotation corpus.

The dataset is described as purpose-built and physics-grounded. This characterization reflects both its generation process and its intended use: rigorous benchmarking of single-image dehazing models in a wildlife domain that differs materially from urban and generic outdoor dehazing benchmarks. A plausible implication is that the dataset’s value lies less in sheer scale than in domain specificity and paired supervision.

2. Source imagery and wildlife domain characteristics

The clear images are selected from the Northeast Tiger and Leopard National Park (NTLNP) dataset, which contains 25,567 original images of 17 species, although AnimalHaze3k uses a subset of 1,159 images representing 11 species (Rai et al., 17 Apr 2026). These source images were captured by infrared camera traps in China’s Northeast Tiger and Leopard National Park between 2014 and 2020. Only daylight images are used; night images are excluded.

The source imagery is characterized by natural, undisturbed habitats with diverse backgrounds, varying weather conditions, and seasonal variations for non-hibernating species. The authors remove timestamp overlays and standardize all images to 640×480 pixels. Sample figures reportedly include animals such as dogs, red foxes, badgers, leopard cats, and Amur tigers, which suggests a mixture of carnivores and other mammals typical of the park, although the paper does not explicitly list all 11 species.

These properties matter because camera-trap wildlife imagery imposes a distinct low-level and mid-level visual regime. Animals are embedded in forest backgrounds, their visibility varies with weather and season, and the imagery is not staged or studio-like. The paper’s framing indicates that this domain specificity motivates both the dataset and the associated model design. This suggests that AnimalHaze3k is intended to fill a gap left by dehazing datasets dominated by urban roads, buildings, and traffic scenes.

3. Physics-based haze synthesis pipeline

AnimalHaze3k is generated through a three-stage pipeline built around the classical atmospheric scattering model (Rai et al., 17 Apr 2026):

$I(x) = J(x)\cdot t(x) + A\cdot(1-t(x))$

with transmission

$t(x) = e^{-\beta \cdot d(x)}.$

Here, $x$ denotes the pixel index, $I(x)$ the observed hazy image intensity, $J(x)$ the clean scene radiance, $A$ the global atmospheric light, $t(x)$ the transmission value, $\beta$ the scattering coefficient, and $d(x)$ the scene depth. In the formulation used by the paper, larger depth yields smaller transmission for fixed $\beta$, while larger $t(x) = e^{-\beta \cdot d(x)}.$0 simulates thicker haze.

The first stage is depth estimation. For each clear image, a metric depth map $t(x) = e^{-\beta \cdot d(x)}.$1 is estimated using HybridDepth, described in the paper as a state-of-the-art single-image metric depth estimation pipeline that fuses depth-from-focus data with relative depth priors. The rationale given for this choice includes higher metric accuracy and better generalization for outdoor scenes, consistency across zoom levels, and the use of Depth Anything as the relative depth estimator.

The second stage is transmission map computation. Given $t(x) = e^{-\beta \cdot d(x)}.$2, transmission is computed as $t(x) = e^{-\beta \cdot d(x)}.$3, with $t(x) = e^{-\beta \cdot d(x)}.$4 drawn uniformly at random from $t(x) = e^{-\beta \cdot d(x)}.$5. The lower end of this range corresponds to lighter haze and the higher end to heavier haze with stronger attenuation. The paper states that the use of a continuous range ensures diverse haze conditions across the dataset.

The third stage is synthetic haze generation. Atmospheric light is sampled from the discrete set

$t(x) = e^{-\beta \cdot d(x)}.$6

Using the scattering equation, each clear image is transformed into three synthetic hazy variants by sampling different $t(x) = e^{-\beta \cdot d(x)}.$7 and/or $t(x) = e^{-\beta \cdot d(x)}.$8. The result is a dataset spanning light to heavy haze, with the variability driven by physically meaningful parameters rather than ad hoc image filters. The paper explicitly positions this as a physics-grounded construction.

4. Benchmark tasks and annotation logic

AnimalHaze3k defines two related evaluation settings: a single-image dehazing benchmark and a downstream detection protocol (Rai et al., 17 Apr 2026). In the dehazing setting, the input is a synthetic hazy wildlife image and the target is the paired clear wildlife image. The prescribed split is the train/validation/test partition already specified above, and the metrics used are SSIM, PSNR, FSIM, and LPIPS.

For downstream detection, the annotation logic changes. The paper states that the authors do not relabel AnimalHaze3k. Instead, YOLOv11 is trained on the original NTLNP dataset using its standard object annotations, including bounding boxes and species labels, and AnimalHaze3k is then used for evaluation. Specifically, hazy test images are passed to YOLOv11, dehazed versions of those same images are also passed to YOLOv11, and performance is compared against ground-truth labels from NTLNP.

This division clarifies a potential misconception. AnimalHaze3k is not presented as a detection dataset with standalone object annotations; it is presented as paired test data that can be linked back to NTLNP annotations for evaluating the effects of haze and dehazing on detection. The dataset’s core supervisory signal is the clear target image, not detection metadata.

The paper further notes that haze condition diversity is implicit rather than stratified into separately reported bins. Because $t(x) = e^{-\beta \cdot d(x)}.$9 and $x$0 is sampled from five discrete values, the benchmark includes a distribution of haze severities, even though results are not reported separately for light and heavy haze. This suggests that generalization across haze density is treated as part of the benchmark rather than as a separate subtask.

5. Coupling with IncepDehazeGan

AnimalHaze3k is tightly coupled with IncepDehazeGan, which the paper introduces as a dehazing model designed for the same wildlife setting (Rai et al., 17 Apr 2026). The stated motivation is that no dedicated hazy wildlife dataset with paired clear-hazy images was available, making it difficult to train and benchmark dehazing models in realistic ecological settings. AnimalHaze3k therefore functions as both the training ground and the evaluation benchmark for the proposed architecture.

The generator is described as an encoder-decoder, U-Net-style network with two key innovations. First, it uses a dual-path encoder. One path consists of standard encoding blocks with two convolution layers, each followed by ReLU activation, and a max pooling layer for spatial downsampling. The second path consists of inception blocks with four parallel convolution layers using kernel sizes $x$1. According to the paper, this multi-scale design is intended to capture both local details and larger context, which is useful because haze affects fine textures and larger regions differently.

Second, the decoder incorporates residual skip connections. Each decoder block contains two standard convolution layers and a transposed convolution for upsampling, while feature maps from each encoder block are fused into the corresponding decoder block. The paper attributes to this design reduced information loss due to downsampling, preservation of low-level details such as edges and textures, and improved gradient flow during training.

The discriminator is a convolutional network resembling a PatchGAN. It has six convolutional blocks, each composed of convolution, leaky ReLU, and batch normalization, followed by a final sigmoid layer that outputs a one-channel matrix rather than a single scalar. The paper states that each value corresponds to a patch’s realism score and that this encourages locally realistic textures while helping avoid overfitting to global appearance.

Training uses a hybrid loss comprising adversarial loss and $x$2 reconstruction loss, with the total generator loss given as a weighted sum and $x$3 in the reported experiments. All models, including IncepDehazeGan and the baselines, are trained on AnimalHaze3k’s training partition with 50 epochs, batch size 4, and an NVIDIA Tesla P100 GPU with 16GB and 3584 CUDA cores.

6. Quantitative performance, ecological role, and limitations

On the AnimalHaze3k test set, the paper compares IncepDehazeGan against FFA-Net, FD-GAN, DEA-Net, and DehazeFormer (Rai et al., 17 Apr 2026). The reported scores for IncepDehazeGan are SSIM 0.8914, PSNR 20.5404, FSIM 0.9363, and LPIPS 0.1104. The corresponding scores for FFA-Net are 0.5468, 12.1842, 0.6665, and 0.3610; for FD-GAN, 0.5580, 17.5573, 0.8314, and 0.1637; for DEA-Net, 0.8303, 18.6481, 0.8936, and 0.2102; and for DehazeFormer, 0.8388, 17.4550, 0.8917, and 0.2375. The paper summarizes this as state-of-the-art performance, including 6.27% higher SSIM than DehazeFormer and 10.2% better PSNR than DEA-Net.

The downstream detection experiment uses YOLOv11 trained on clear NTLNP images and evaluated on AnimalHaze3k test images before and after dehazing. The reported hazy-image performance is mAP 0.3216 and mIoU 0.4313. After dehazing, the reported performance is mAP 0.6842 and mIoU 0.7201. The paper interprets this as an approximately 112% relative improvement in mAP and approximately 67% relative improvement in mIoU.

These results are situated within a conservation-oriented use case. The paper states that haze degrades contrast between animals and background, visibility of anatomical features such as stripes and spots, and depth cues necessary for behavior analysis. It identifies affected downstream tasks as species identification, individual re-identification, population counting, movement and behavior tracking, and poaching and illegal activity surveillance. In this framing, AnimalHaze3k acts as a bridge between low-level image enhancement and high-level ecological analytics.

The main limitations are also explicit. AnimalHaze3k uses synthetic haze generated from estimated depth maps, so a synthetic-versus-real haze domain gap remains. The dataset uses 11 species from a single region, only daylight infrared camera-trap images, and a first-order atmospheric scattering model with exponential transmission and global atmospheric light. Evaluation is reported only on synthetic haze. The paper suggests future directions including collecting real hazy wildlife imagery when feasible, performing domain adaptation between synthetic and real haze, expanding to more species and ecosystems, extending to nighttime or thermal images and other adverse conditions, and exploring more advanced physics-based atmospheric models. Despite these limitations, the paper concludes that synthetic, physics-based haze is sufficient to train a dehazing model with strong quantitative metrics and substantial gains in downstream animal detection performance.